Category: 3. Business

Introduction

Idiopathic normal pressure hydrocephalus (iNPH) is a neurological disease characterized by cognitive impairment, gait disturbance, and urinary incontinence [], with ventricular enlargement and the DESH sign (disproportionately enlarged subarachnoid hydrocephalus) shown on brain scans []. It is a common cause of dementia in older people [] and can be treated effectively with cerebrospinal fluid (CSF) shunts [,].

However, the clinical manifestations of iNPH lack specificity, with only 50% of patients exhibiting the full triad []. Therefore, patients with iNPH are usually misdiagnosed as other neurodegenerative diseases, such as Alzheimer disease or Parkinson disease []. Unlike these conditions, the clinical symptoms of iNPH can be reversed through CSF shunting []. However, only 39% to 81% of patients with shunts show improvement 3 to 6 months after the shunt insertion []. The CSF drainage test is an important tool for evaluating patients with possible iNPH before shunt surgery [,,]. The tap test and continuous external lumbar drainage (ELD) are 2 commonly used methods []. Previous studies have shown that the tap test has a sensitivity of only 26%, whereas the continuous ELD has a sensitivity of 50% and a specificity of 80% [,]. As a result, some medical centers conduct continuous ELD directly to avoid repeated lumbar punctures and reduce the risk of misdiagnosis [-]. On the other hand, current evaluation methods for CSF drainage tests also have relatively low sensitivities. Previous studies using the Mini-Mental State Examination (MMSE), 10-m walking test (TMWT), and timed up and go (TUG) test indicate a sensitivity of 26% for tap test [,]. This implies that current assessment methods are inadequate in providing reliable guidance for clinical practice because traditional cognitive tests can cause large learning effects because of the short time intervals between tests [] and video-based gait analysis may be influenced by the subjective opinions of raters [] and lacks quantitative measures, such as stride length, step height, and gait velocity [,]. These shortcomings of traditional tests highlight the need for a more objective and quantitative evaluation method.

In recent years, digital neuropsychological evaluation equipment has emerged, which was designed to overcome the major limitations of paper and pencil tests, such as low sensitivity, subjectivity, and test-retest reliability []. They also provide randomized test paradigms and automated recording of variables such as reaction time, thereby improving evaluation efficacy []. A recent study indicated that computerized neuropsychological tests could detect cognitive impairment and postshunt improvements in patients with iNPH []. Additionally, 3-dimensional gait analysis aids in the diagnosis of iNPH by offering multidimensional gait parameters [,]. There is growing evidence demonstrating that motor and cognitive impairment in iNPH may result from interconnected neural network impairments [,], and recent research based on traditional cognitive and gait tests found that the improvement of different symptoms after ELD may follow different temporal trajectories []. Therefore, combining cognitive and gait assessments may provide a more comprehensive and complementary picture of functional response after ELD. Within the expanding field of digital medicine, the integration of digital cognitive and gait tests into diagnostic workflow has been widely investigated and implemented in other cognitive disorders [,]. However, an evidence gap remains regarding the predictive value of these digital tests in the context of ELD in patients with iNPH.

Therefore, this study aimed to (1) evaluate the improvement of cognitive and gait parameters after ELD through the application of digital tests and (2) investigate the predictive value of digital cognitive and motor assessments for shunt outcomes in patients with iNPH.

Methods

Study Design

Patients with iNPH were enrolled from an ongoing prospective cohort study at West China Hospital of Sichuan University between May 2022 and November 2023. The inclusion criteria included (1) at least one symptom of cognitive decline, gait disturbance, and urinary incontinence; (2) ventricular enlargement (Evans index >0.3), focally dilated sulci, or the DESH sign shown on brain magnetic resonance imaging; (3) CSF opening pressure ≤200 mm H₂O; and (4) informed consent. We excluded patients with (1) gait disturbance, cognitive impairment, or urinary incontinence due to other neurological diseases (cerebral hemorrhage, brain trauma, brain tumor, and intracranial infection); (2) inability to complete quantitative motor or neuropsychological tests; and (3) refusal to undergo or a negative response to continuous ELD. After enrollment, all participants underwent comprehensive baseline clinical assessments and continuous ELD. Detailed methodologies were described in the following sections. The participant flow throughout the study is summarized in . This cohort study was conducted in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) reporting guidelines (Checklist 1).

Ethical Considerations

This study was approved by the Ethics Committee of the West China Hospital, Sichuan University (No. 2022‐538). To ensure confidentiality, all data were anonymized before analysis and stored securely, with access restricted to the research team. Participants were not compensated financially but volunteered after being informed of the study’s potential benefits.

Demographic Characteristics

We systematically collected demographic information of the enrolled patients, including age, sex, years of education, medical history, and conducted a comprehensive neurological examination. The Japanese iNPH grading scale (iNPHGS) was used to quantify the severity of symptoms of iNPH [,]. Each symptom was scored from 0 to 4, with a total score of 12; higher scores indicating more severe clinical symptoms. The modified Rankin Scale (mRS) was used to assess the patients’ daily living function [], scored from 0 to 5, with higher scores indicating a more severe inability to perform daily living.

Neuropsychological Assessments

All neuropsychological tests were conducted by 2 trained neuropsychologists. Traditional neuropsychological assessment uses the MMSE to evaluate global cognitive function []. The BrainFit computerized neuropsychological assessment device (Beijing CAS-Ruiyi Information Technology Co, Ltd, China) was used to provide computerized neuropsychological tests, including the grammatical reasoning test, the one-back test, the trail-making test (TMT), and the Stroop color-word test (SCWT). The reliability of the BrainFit system was validated among Chinese populations in previous studies [,]. The grammatical reasoning test is designed to evaluate the language comprehension and reasoning ability of the participants []. In 60 seconds, they had to determine if the way the shapes were arranged on the screen matched the textual description. The shapes were a circle and a square with a particular positional relationship, and in the middle of the screen was a sentence that described this relationship. If the sentence and shape combination matched, participants clicked “correct”; if not, they clicked “incorrect.” A new question would display within a second of each click []. The score for this test was calculated as the number of correct responses minus the number of errors. The one-back test is based on the “N-back” working memory paradigm []. Within 60 seconds, participants were required to click the “same” or “different” button to indicate whether they thought the current displayed number matched the previous one. A new number appeared within 1 second after each click. The score was calculated as “the number of correct responses minus the number of incorrect responses” []. The Trail Making Test primarily assessed the participant’s attention and reaction speed, requiring the participant to sequentially connect 25 randomly arranged numbers on a page, with the test time recorded []. The Stroop Color-Word Test was used to evaluate executive function, psychomotor speed, and cognitive flexibility. During this study, participants were asked to quickly name the color of words that describe different colors, with test time and the number of correct responses recorded [].

Gait Analysis

Traditional gait analysis includes the TMWT and the 5-m TUG test [,]. The TMWT requires the patient to walk on a 10-m-long path while recording a video. The total time and the number of steps were recorded. The TUG test requires the patient to walk on a 5-m path from one end to the other, turn 180 degrees, and then walk back to the beginning. Each test recorded the overall time, number of steps, and number of turns necessary. Each of these tests was repeated 3 times, and the average of the 3 results was recorded.

Quantitative gait analysis was evaluated using a 3-dimensional gait analysis system ReadyGo (Beijing CAS-Ruiyi Information Technology Co, Ltd), the accuracy and sensitivity of the system have been validated in previous studies [,,]. The ReadyGo system uses a single-camera setup to record 3-dimensional motion by using deep learning for precise positioning of skeletal points. For the single-gait test, participants were instructed to walk at their habitual pace on a 3 m walkway. Step width, stride length, step height, gait velocity, and turning time were the analyzed parameters. Step width was defined as the average width between the left and right feet in each image frame. A single-foot stride length was the distance between 2 landings from the same foot; the overall stride length was the average of the left and right sides. The step height was the height at which a foot could swing without touching the ground. The average step heights of the left and right feet were used for the final analysis. Gait velocity was defined as the distance between the start and end points divided by test time [].

Continuous ELD Test

Patients with possible iNPH underwent lumbar puncture and continuous ELD after obtaining informed consent. For 3 to 5 days, the daily CSF drainage volume was controlled between 100 and 150 mL [,,,]. The patients underwent traditional cognitive and gait assessments at baseline and every day after ELD, with extra digital neuropsychological and gait assessments conducted at baseline and on the third day following ELD. The criteria for a positive response to the traditional evaluation method of ELD included (1) improvement of ≥20% in either time or number of steps in the 10-m walking test or ≥10% improvement in both time and number of steps, (2) improvement of ≥10% in the TUG test time, or (3) improvement of ≥3 points in the MMSE score [,,].

Postoperative Follow-Up

Patients diagnosed with probable iNPH after ELD underwent lumboperitoneal shunt placement after informed consent was obtained. These patients were followed up regularly at 3 months, 6 months, and 1 year postoperatively. Follow-up was conducted through in-person visits or telephone interviews with the patients and their long-term caregivers. During the in-person visits, comprehensive neurological examinations and brain imaging were performed. Objective assessments of cognitive function, gait, urinary function, daily living abilities, subjective symptom improvements reported by the patients and caregivers, and postoperative complications were documented. Telephone interviews primarily assessed whether the patients and caregivers reported symptom improvement and identified any possible complications. Patients who showed an improvement of ≥1 point in mRS score compared to baseline or an improvement of ≥1 point in any iNPHGS score were defined as definite iNPH (shunt responders) [,]. Those who did not report symptom improvements during the follow-up period were classified as nonresponders and underwent pressure readjustment and long-term follow-up.

Statistical Analyses

The raw scores of grammatical reasoning, the one-back test, TMT, SCWT correct number, and reaction time were converted into standardized Z-scores based on previously reported norms for the Chinese population, facilitating comparisons across tests [,]. The z-scores for the Stroop C correct number, grammatical reasoning test, and one-back test were calculated using the formula (raw score−mean)/SD. Z-scores for TMT and Stroop C reaction time were derived by subtracting the raw score from the mean and dividing by the SD. Lower z-scores indicated more significant impairments in the corresponding cognitive domain. The cognitive Z-score for every patient was calculated by taking the mean of the Z-scores from the tests mentioned earlier [,]. The improvement rate of the cognitive Z-score was calculated as follows: (post-ELD parameter−pre-ELD parameter)/pre-ELD parameter×100%. For quantitative gait analysis, the improvement rates in gait velocity, stride length, and step height post-ELD were calculated as (post-ELD parameter − pre-ELD parameter)/pre-ELD parameter×100%. The improvement rates of the step width and turning time post-ELD were calculated as follows: (pre-ELD parameter−post-ELD parameter)/pre-ELD parameter×100%. The gait improvement rate was calculated as averages of all 5 gait parameters, and the combined improvement rate was averaged based on the improvement rate of the cognitive Z-score and gait improvement rate []. In this pilot study, gait and cognitive improvement were assigned equal weight in the combined improvement rate to reflect global functional improvement after ELD, as the gait and cognitive impairment are core symptoms in iNPH and are considered equally important in current clinical assessments such as iNPHGS [,].

For continuous variables, normally distributed data are shown as mean (SD), whereas nonnormally distributed data are presented as median (IQR). Group comparisons were performed using the Student t test for normally distributed data and the Mann-Whitney U test for skewed data, whereas comparisons of proportions were performed using the χ² test. Group comparisons before and after the ELD were performed using paired t tests for normally distributed data and Wilcoxon signed-rank tests for nonnormally distributed data. To address the limited sample size, we used Firth penalized logistic regression to evaluate the predictive value of improvement rates from each evaluation method for shunt response, and variables demonstrating significant difference in group comparison were included in subsequent multivariate models. This method effectively reduced the small-sample bias and provided more reliable coefficient estimates []. The results of Firth logistic regression models were reported by odds ratios (ORs) and receiver operating characteristic curves. Optimal cutoff values were determined based on the Youden index, and diagnostic metrics, including area under the curve (AUC), sensitivity, specificity, positive predictive value, and negative predictive value, were calculated. The DeLong test was used for AUC comparison between traditional tests and digital tests. Permutation tests with 5000 iterations on all Firth regression models and internal validation of the diagnostic metrics with 2000 bootstrap resamples were used to assess the model performance and further enhance the robustness of the statistical inference []. A P value of .05 was considered statistically significant. All statistical analyses were performed using STATA SE 16.0 (StataCorp, TX, USA) and R version 4.3.0 (R Foundation).

Results

Demographic and Clinical Characteristics of Patients With Probable iNPH

A total of 127 patients diagnosed with possible NPH were consecutively enrolled in this study (). After excluding 7 patients who refused lumbar puncture, 13 patients who developed secondary NPH, and 30 patients who were unable to cooperate with the quantitative motor and cognitive tests, 77 patients with possible iNPH have finally received continuous ELD. Of them, 70 patients were ELD responders and were diagnosed as probable iNPH by clinical neurologists, whereas 7 were negative for ELD tests. Among the 70 patients with probable iNPH, 31 refused shunt surgery and were ultimately diagnosed with probable iNPH, and 39 (55.7%) underwent lumboperitoneal shunt surgery.

‎

presents the baseline characteristics of the 70 patients with probable iNPH. The median (IQR) age was 74 (69-80) years, and 70% (49/70) of the patients were male. The prevalence of cognitive impairment, gait disturbances, and urinary incontinence was 92.86% (65/70), 95.71% (67/70), and 54.29% (38/70), respectively, with 51.43% presenting with the full Hakim triad (). We compared the baseline characteristics between patients with iNPH who received shunt surgery and those who did not, and the shunted group exhibited a significantly more severe DESH score on brain magnetic resonance imaging (P=.007), whereas the 2 groups did not differ in demographics, symptom severity, or improvement rates after ELD. This suggests that the 2 groups were largely comparable at baseline.

As shown in and in , among the 70 patients with probable iNPH, traditional tests, including MMSE (P<.001), TUG (P=.004), TMWT time (P=.016), and steps (P<.001), showed significant improvement 3 days after ELD. For digital tests, one-back test (P=.01), SCWT correct numbers (P=.009), and composite cognitive z-scores (P=.025) improved significantly 3 days after ELD. Quantitative gait analyses also showed significant improvement in stride length (P=.01), step height (P=.048), gait velocity (P=.019), and reduced turning time (P<.001) after 3-day ELD compared to baseline.

‎

Predictive Value of Digital Neuropsychological and Gait Tests for Shunt Outcome in Patients With iNPH

A total of 39 patients with probable iNPH underwent lumboperitoneal shunt placement; after a median follow-up interval of 116 days, 34 (89.7%) showed at least a 1-point improvement in mRS or iNPHGS, leading to a clinical diagnosis of definite iNPH or being classified as “shunt responders,” 4 patients only reported subjective symptomatic improvement but not measurable through mRS or iNPHGS, and 1 patient did not report any improvement at the time of final follow-up and was managed with pressure readjustment and continued follow-up.

First, we compared the clinical characteristics between shunt responders and nonresponders (). The results demonstrated a higher proportion of male sex (P=.036), lower Evan index (P=.035), and significantly higher improvement rate of gait after ELD (P=.04) for shunt responders, whereas they did not differ significantly in age, educational level, or vascular risk factors. Multivariate Firth logistic regression model adjusted for sex and Evan index showed a significant association between a higher improvement rate of digital gait analysis and a lower risk of unfavorable shunt outcome (adjusted OR 0.90, 95% CI 0.78‐0.99; P=.03). In addition, an association between a higher combined improvement rate of digital neuropsychological and gait analysis and a lower risk of unfavorable shunt outcome was also observed (adjusted OR 0.98, 95% CI 0.95‐1.00; P=.03; ).

As shown in and , traditional tests in ELD yielded poor diagnostic performance, with an AUC of 0.55 (95% CI 0.30‐0.81), sensitivity of 40% (95% CI 12%‐77%), specificity of 71% (95% CI 54%‐83%), PPV of 17% (95% CI 5%‐45%), and NPV of 89% (95% CI 72%‐96%). In contrast, the combined digital cognitive and gait approach yielded an AUC of 0.92 (95% CI 0.83‐1.00), sensitivity of 100% (95% CI 57%‐100%), specificity of 79% (95% CI 63%‐90%), PPV of 42% (95% CI 19%‐68%), and NPV of 100% (95% CI 88%‐100%). The bootstrap-derived cutoffs showed moderate dispersion, whereas the predictive performances of all digital tests remain relatively stable (lower limits of all AUCs >0.81). The DeLong test demonstrated that combining digital cognitive and gait tests showed better predictive performance of shunt outcome compared to traditional tests (Z=2.43; P=.015).

Furthermore, the calibration curve demonstrated that the combined improvement model showed good overall calibration quality (, Spiegelhalter P=.62), with a low average calibration error of 3.0%, despite slight overconfidence (calibration slope=1.299).

‎

Discussion

Principal Findings

In this study, we investigated the predictive value of digital neuropsychological and gait analyses during ELD for shunt outcome in patients with iNPH. Our findings revealed that while both traditional and digital cognitive and gait assessments improved significantly after 3 days of ELD, the digital tests outperformed traditional testing in terms of predicting shunt outcomes.

Our study found that patients with probable iNPH showed gait improvement after ELD, which was primarily manifested as increased gait speed, longer stride length, higher step height, and shorter turning time. This is consistent with the findings of prior investigations using quantitative gait analysis []. Electronic walkways, wearable sensors, and accelerometers are all commonly used for quantitative gait analysis [,]. A prospective study using electronic walkways discovered that patients with iNPH had increased stride length and gait speed following a tap test []. Another research study has revealed that 72 hours following the tap test, patients with probable iNPH had longer stride lengths, shorter double support times, and faster cadences []. A recent study based on 3-dimensional gait analysis indicated that patients with probable iNPH may show improvements in spatiotemporal parameters, such as step length, gait speed, and cadence 24 hours after the tap test, with improvement rates ranging from 4.9% to 10.5% []. Besides, a preliminary study using a video-based method qualitatively assessed gait changes after the tap test and found significant improvements in step height [], whereas earlier studies based on inertial sensors and electronic walkways may not provide precise quantitative assessments of step height. Our vision-based system provided richer spatiotemporal gait parameters than previous methods, without the need for wearable sensors. Therefore, it can provide improvements in step height following ELD, which partially complements the findings of previous studies.

Cognitive impairment is the second most common symptom in patients with iNPH, which may be partly due to mechanical stress to the brain and the existence of Alzheimer disease co-pathologies [,]. Frontal lobe dysfunction is thought to be the classic cognitive profile in patients with iNPH []. In our study, we observed significant improvements in one-back tests and Stroop color-word tests post-ELD, indicating improvements in executive function, attention, and working memory, all of which are associated with improved frontal lobe functions. Patients with iNPH often show significant improvements in executive subfunction after CSF drainage [], with prior research showing improvements in verbal fluency, frontal assessment battery, trail-making test, and Stroop color-word tests [-]. Therefore, executive subfunction assessment is emphasized in evaluations during CSF drainage to distinguish iNPH from its mimics [,]. Previous studies have developed an executive function battery for patients with iNPH based on tests such as TMT-A, Stroop color-word tests, and digit symbol substitution tests [], and this battery demonstrates a sensitivity of 80% in predicting shunt responders with a specificity of 100%, indicating that evaluating executive function, attention, and reaction time before and after CSF drainage may be crucial in identifying shunt responders [].

Despite the widespread use of the CSF drainage test for diagnosing and predicting shunt prognosis in patients with iNPH, earlier research indicated that the sensitivity of a single-tap test might be as low as 26%, whereas continuous ELD had a sensitivity of approximately 50% [,]. Although the tap test response would be influenced by morphological features and CSF dynamic changes [], the method used to evaluate the CSF drainage test itself may be crucial for determining its overall sensitivity [,]. In the latest Japanese iNPH management guidelines, MMSE, TMWT, and TUG tests were recommended to assess changes in patients’ gait and cognitive function []. However, using these methods for assessment may result in very low sensitivity and negative predictive value []. A previous large-scale European multicenter study indicated that the negative predictive value of a single lumbar puncture drainage could be as low as 18% [], with other literature reports generally ranging from 18% to 50% []. In our cohort, traditional cognitive and gait tests exhibited a sensitivity of only 40% and a specificity of 70%, which demonstrated that traditional tests may potentially lead to misdiagnosis of shunt responders. Therefore, traditional assessment methods may be insufficient to guide the selection of shunt candidates [].

In recent years, digital neuropsychological evaluation equipment has emerged that provides randomized test paradigms and automated recording of variables such as reaction time, thereby improving test efficacy []. Preliminary research has used digital cognitive and gait evaluation tools for patients with iNPH, establishing computerized neuropsychological tests as reliable techniques for diagnosing cognitive impairments and postoperative cognitive improvements in patients with iNPH []. In contrast, a recent meta-analysis suggested that quantitative gait analysis can detect gait improvements in patients with iNPH at baseline, after the tap test, and postshunt []. These sensor-based technologies and digital systems possess high scalability, offering continuous and high-precision monitoring of cognitive and motor functions []. Our results further emphasized the predictive value of these digital assessment tools during ELD for patients with iNPH, which also showed better performance in predicting shunt responses. Therefore, the use of digital evaluation tools may aid in the selection of potential shunt candidates. Integrating these digital data into telehealth platforms would enable remote and continuous health care for patients with geographic barriers, reduce hospital visits, and enhance patient outcomes []. However, the results of our study are still preliminary; research gaps for digital assessments still exist in modest sample size, clinical standardization, and multicenter validation. Future research should explore more data-driven weighting schemes and conduct external validation as more data become available.

Strengths and Limitations

The main advantages of our study are the prospective cohort study design, which collected detailed clinical data, and comprehensive neuropsychological and gait assessment data from the patients. Additionally, we are the first to combine computerized neuropsychological assessment and 3-dimensional gait analysis to evaluate symptom changes during ELD in patients with iNPH. The limitations of the study are as follows: (1) Nearly half of the patients with positive ELD responses refused shunt surgery, and the relatively small sample size of shunted patients may constrain the statistical power, despite our use of Firth penalized regression models. Therefore, the relatively high AUC and sensitivity values should be interpreted with caution. Future multi-center studies with larger samples are needed to validate our preliminary findings. (2) Cut-off values were calculated statistically based on the Youden index. Given the relatively small sample size, these data-derived thresholds should be interpreted as preliminary estimations that require external validation in future studies with larger cohorts. (3) The high rate of shunt refusal may inevitably introduce selection bias, although the baseline characteristics between patients who accepted and refused shunts were largely comparable, except for DESH score (). (4) Several patients were excluded because of severe illness; therefore, these results may be primarily applicable to patients who have iNPH with mild-to-moderate symptoms. (5) In this study, quantitative assessments were only completed on the third day after ELD. Although previous studies indicate that a 3-day continuous ELD is sufficient to improve cognitive and gait function in patients with iNPH [,], a recent longitudinal study suggests that patients with iNPH would exhibit delayed response up to 2 weeks after ELD []. Therefore, further research is needed to determine the best evaluation time point. (6) Due to the relatively short follow-up interval, the long-term predictive value of these digital tests for shunt response was unclear, which may be affected by various prognostic factors []. (7) Although our findings on digital assessments are promising, the generalizability across different health systems and the cost-effectiveness of these digital assessments are underexplored.

Conclusions

Digital neuropsychological and motor assessments could identify gait and cognitive improvements after continuous ELD. Moreover, the digital tests showed better predictive performance for shunt outcome compared to traditional tests in our pilot study. Further validation of our findings in a multicenter setting is required to establish the optimal cutoff values for these digital evaluation tests.

Introduction

There is a critical need for scalable, cost-effective interventions to address high rates of physical inactivity and related chronic diseases, especially in underserved populations considered to be at risk. In fact, the World Health Organization estimates that further inaction could lead to 500 million new global cases of preventable noncommunicable diseases, with associated direct health care costs of US $520 billion, by 2030 []. Internet-based physical activity (PA) interventions have great potential for widespread dissemination and have already been shown to improve PA levels [-]. However, most of this work was conducted in the general population and not extended to those most at risk, including underrepresented minority populations.

Hispanic adults report the highest rates of inactivity outside of work (32.1% vs 23% for non-Hispanic White individuals) and experience disproportionate rates of associated conditions (eg, obesity and diabetes) [,]. In fact, 79% of Latinas are overweight or obese, compared to 64% of non-Hispanic White women []. Thus, in the Pasos Hacia La Salud study, we tested a web-based, PA intervention in Spanish with 205 Latinas and found significantly greater increase in weekly minutes of moderate-to-vigorous PA (MVPA) compared to a control condition [,]. Moreover, the costs associated with providing the intervention were low at 6 months (US $17 per person per month) and further decreased (US $12 per person per month) at 12 months []. While the Pasos program was shown to be effective and low cost, most intervention participants (69%) did not meet national PA guidelines at 6 months [,,], and longer-term maintenance was not evaluated.

To help sedentary Latinas achieve and maintain health-enhancing levels of PA, the intervention was refined by adding SMS text messages and further targeting key social cognitive theory [] variables (eg, self-efficacy, enjoyment, and social support). Subsequently, in the Pasos Hacia La Salud II study, the new technology- and theory-enhanced version was tested and compared with the original Pasos Hacia La Salud intervention in a randomized controlled trial (N=205) []. There were no significant between-group differences in PA at 6 or 12 months; however, the enhanced intervention arm had higher PA levels than the original intervention arm at 18 and 24 months [].

While both programs have been shown to be beneficial, the enhanced intervention appears to have an advantage in terms of long-term behavior change. However, these enhancements (eg, SMS text messages, discussion board, and additional staff contacts) likely come at additional expense (compared to the original arm). An evaluation of intervention costs is needed to determine which program is feasible and the best fit for the clinical and community setting and resources. Thus, to inform future dissemination and implementation efforts, this study examined additional costs associated with intervention enhancements and how they influence cost-effectiveness.

Methods

Overview

Pasos Hacia La Salud II was a fully powered randomized trial of 2 PA interventions: the original web-based PA intervention (original) and an enhanced arm that included additional elements to support increasing activity (enhanced). Participants were adult Latinas aged 18 to 65 years who were underactive (engaging in <60 min of MVPA per week), could read and write in Spanish, had access to technology to use the study website and receive SMS text messages, and were healthy enough for unsupervised exercise. A full description of the study protocol, measures, and participants has been published previously [,].

Intervention Description

In the original arm of the web-based intervention, participants attended a baseline session led by trained bilingual interventionists who explained MVPA and helped participants set incremental goals. Participants also completed goal-setting sessions over the phone at 1 and 9 months and at follow-up visits (6 and 12 mo). Participants were encouraged to accumulate MVPA on most days of the week and encouraged to achieve the goal of 150 minutes of MVPA by 6 months. All participants were given a pedometer (Accusplit Inc) and were encouraged to use it to self-monitor daily activity.

The interventionist also oriented participants to the study website, which participants could access throughout the intervention period (24 months). The website included tools to help participants increase their MVPA, such as self-monitoring, goal setting, social support, instruction about how to be physically active, problem solving, motivational stage-matched PA manuals, and individually tailored feedback. Further details about the tailored intervention content are described elsewhere []. Participants were encouraged to regularly access the website, log their PA on the website, and receive a tip of the week. Participants were also prompted to complete online surveys to customize their intervention content. Most web-based content was automated and did not require staff time. However, staff regularly checked the message boards to answer questions as necessary and were available to help participants with any technology issues.

In addition to the intervention components described earlier, the enhanced intervention included (1) SMS text messages; (2) additional phone calls at 2, 3, 15, and 21 months; (3) additional in-person visits at 18 and 24 months; and (4) additional data-driven content and website features that encouraged user participation. Participants were also provided with nonmonetary incentives for their website use. Participants earned points by logging into the website and using certain features, such as the goal-setting calendar. Once participants earned a given number of points, they were sent prizes, such as water bottles, phone cases, and T-shirts.

To further enhance social support for PA, a discussion board was used to facilitate meetups among participants. Research staff posted details about free and low-cost PA events in the area so that participants could meet, discuss coordinating attendance with others, overcome barriers to PA, and motivate each other to attend these events.

Ethical Considerations

The study was reviewed and approved by the Brown University Human Research Protections Program (1708001868), and all participants provided written informed consent in the language of their choosing (English or Spanish). To protect confidentiality, all data were deidentified for storage and analysis. To thank them for their time, participants received compensation for participating in study activities. Participants received US $25 for completing each of their assessment visits (at baseline and 6, 12, 18, and 24 months) with an additional US $50 bonus at 24 months if they completed all visits. In addition, all participants received incentives for filling out monthly questionnaires (US $10 per month) to generate the tailored website content. Participants also received reimbursement costs related to travel and childcare costs. Participants in the enhanced arm also received prizes for engaging with website content (as mentioned earlier). This trial has been registered at ClinicalTrials.gov (NCT03491592),

PA Measures

The main outcomes used to determine cost-effectiveness were the change in minutes of MVPA from baseline to 12 and 24 months, as measured by ActiGraph accelerometers and self-report, and the percentage of participants meeting national guidelines at 12 and 24 months. Self-reported activity was measured using the 7-day PA recall (PAR) interview, an interviewer-administered measure that has shown good reliability and validity in adult populations, including in Spanish populations [-]. Adherence to national guidelines was defined as engaging in ≥150 minutes per week of MVPA, as measured by the 7-day PAR [].

Costs

Overview

Cost calculations followed the model used in our previous publications []. Costs were calculated from a payer perspective to estimate the amount it would cost to deliver the intervention in a clinical or community setting. This included all costs associated with personnel, materials, and use and maintenance of technology to deliver the intervention, sourced from actual costs incurred during the trial.

Costs associated with research, such as recruitment, baseline and follow-up measurement visits, obtaining consent, or compensation for research participation, were not included. We also did not include costs associated with intervention development. Thus, both research and development costs are not relevant to the future implementation of the existing intervention in community or clinical settings. Excluding these costs is consistent with published guidelines on cost-effectiveness analyses [].

Personnel Time

Staff time logs were used to determine the time needed to complete each activity associated with intervention delivery. This included training time (for both the trainer and the trainee); time to compile study materials; time for study visits and phone calls (depending on study condition); and time for study maintenance activities, such as checking the online message board, resending pedometers to participants who lost them, calling the website developer with technical questions, or (in the enhanced group) offering SMS text message support. Study maintenance activities were conducted during a set time each week, rather than calculated as a per-participant activity.

Costs for personnel were calculated by determining the total time needed to deliver the intervention per participant in each condition and multiplying by actual staff salary rates, including benefits. Staff delivering the intervention were entry-level research staff with an undergraduate degree (earning a yearly salary of US $54,080 plus 31% in benefits, for a total hourly cost of US $34.05). Training staff were master’s-level behavioral scientists (earning a yearly salary of US $68,640, plus 31% in benefits, resulting in a total hourly cost of US $43.20). Overhead was estimated as an additional 10% of all personnel costs to account for the use of shared space.

Technology

Web hosting was US $75 per month for both study arms. The automated SMS text messaging was supported through the study website and thus did not incur any additional costs.

Materials

Costs for materials were based on actual costs incurred during intervention delivery. All participants in both study arms received an Acusplit pedometer to record their daily steps for US $20 each. Participants in both arms also received a study binder at their baseline visit with information about exercise, tip sheets, and logs for using their pedometers. The total cost for the binders and printed materials was US $4.95 for each participant.

Total cost for prizes in the enhanced group was calculated based on the actual cost of the items and the number of each item that was distributed. As not all participants earned all prizes, we calculated an average cost per participant for prizes to allow for an estimate of the actual cost of delivering prizes at varying rates based on participant engagement. Prize costs ranged from US $1.11 each for water bottles to US $12.50 for a yoga mat. Total cost for prizes in the enhanced arm was US $2038.

A computer was also purchased for the study staff to conduct baseline study visits, check the study website message boards, etc. The cost of the computer was US $1200.

Analysis

To allow for comparison, we used the same analysis approach as in the parent study []. Costs were calculated from a payer perspective to estimate the cost of implementing the intervention in a community or clinical setting. Total costs were calculated separately for each study arm and included the total of personnel time (including benefits and use of shared space), materials, and web hosting. Total costs for each arm were divided by the number of participants in each study arm to calculate the cost per person to deliver each intervention. Costs were calculated in US dollars for 2024.

Cost-effectiveness was calculated as the cost per additional minute of MVPA per person over the course of the study. Total increase in minutes was calculated using linear interpolation of minutes across measurement time points (baseline and 6, 12, 18, and 24 months) and subtracting baseline minutes. Main outcomes for costs and cost-effectiveness were calculated at 12 months (main study outcome) and cumulatively at 24 months. The cost of intervention delivery per person was divided by the cumulative increased minutes per person over the course of the study to estimate the cost per additional minute per person. This was done both for self-reported activity (7-day PAR) and objectively measured activity (ActiGraph accelerometer). Similarly, the cost of each intervention was divided by the number of participants in that group who met national guidelines at each time point to estimate the cost of moving 1 person from being inactive to successfully meeting guidelines for activity at 12 and 24 months.

Incremental cost-effectiveness ratios (ICERs) were calculated to determine the additional cost per minute of MVPA in the enhanced group beyond that achieved by the original group, as well as the cost per additional individual meeting guidelines in the enhanced group compared with the original group. ICERs were calculated by dividing the difference in the costs between the 2 study arms by the difference in cumulative minutes between the 2 arms and, separately, by the difference in the number of people meeting guidelines. For ICERs for guidelines, CIs were computed using the nonparametric bootstrap with 1000 replications.

Results

Overview

Participants (N=195) had a mean age of 43.3 (SD 10.29) years and were primarily of Dominican (80/195, 41%) and Colombian (33/195, 16.9%) descent. Approximately half of the participants (92/195, 47.2%) had a high school education or less. A full description of baseline characteristics and a CONSORT (Consolidated Standards of Reporting Trials) diagram have been published previously [].

PA Changes

As published previously [], participants in the enhanced group (103/195, 52.8%) increased self-reported weekly MVPA from 57.9 minutes at baseline to 115 minutes at 6 months, 106.8 minutes at 12 months, and 135.3 minutes at 24 months, while the original group (92/195, 47.2%) increased self-reported weekly MVPA from 55.2 minutes at baseline to 88 minutes at 6 months, 111.7 minutes at 12 months, and 93.9 minutes at 24 months []. ActiGraph-measured weekly MVPA in the enhanced group changed from 19.7 minutes at baseline to 47 minutes at 6 months, 44.5 minutes at 12 months, and 47.4 minutes at 24 months. In the original group, ActiGraph-measured weekly MVPA changed from 20.6 minutes at baseline to 43 minutes at 6 months, 55.9 minutes at 12 months, and 31.2 minutes at 24 months.

None of the participants met national MVPA guidelines at baseline. At 12 months, 28.2% (29/103) and 29% (27/92) of participants met guidelines according to self-report in the enhanced and original conditions, respectively. At 24 months, this increased to 36.9% (38/103) and 31% (29/92) in the enhanced and original groups, respectively.

Costs

Total costs associated with delivering the interventions are shown in . Total cost of delivering the original intervention at 12 months was US $15,741, or US $171 per person (US $14 per person per month). The enhanced intervention cost US $20,435 to deliver over the first 12 months, or US $198 per person (US $16 per person per month).

At 24 months, cumulative costs for the original intervention increased slightly to US $17,428, or US $189 per person (US $8 per person per month). However, the cumulative costs for the enhanced intervention increased to US $30,862 at 24 months, or US $300 per person (US $12 per person per month). The largest source of the difference was personnel cost for intervention delivery between 12 and 24 months ( and ).

Cost-Effectiveness

Cost per increased minute of activity is shown in . For self-reported activity, each additional minute gained over 12 months in the enhanced group cost US $0.09, compared with US $0.11 per minute in the original group. Incremental costs of increased minutes in the enhanced group were US $0.05 per minute beyond those reported by the original group.

As increases in MVPA measured by the ActiGraph were smaller, these also cost more, with each increased minute costing US $0.19 in the enhanced group and US $0.16 in the original group. As increases in the original group were larger than those in the enhanced group, ICERs could not be calculated.

Costs for cumulative increases over the 24-month period were lower. Each self-reported increased minute cost US $0.06 in the enhanced group and US $0.05 in the original group. Incremental increases in the enhanced group beyond the original group were US $0.08 per additional minute of self-reported MVPA. ActiGraph-recorded minutes over the 24-month period were US $0.12 in the enhanced group compared to US $0.10 in the original group, with incremental minutes costing US $0.20 each in the enhanced group beyond that in the original group.

Cost per person meeting guidelines was normalized per 100 people in each arm. This was higher in the enhanced arm at 12 months (US $705) than in the original arm (US $503). These rose to US $812 and US $601 at 24 months, respectively. ICERs at 24 months were US $1837 (95% CI US $730.89-US $2673.89) per additional person meeting guidelines in the enhanced arm beyond those in the original arm.

Discussion

Principal Findings

Analyses showed that the technology- and theory-enhanced PA intervention for Latinas was more costly than the original intervention but still markedly less costly than most medical interventions. The enhanced intervention cost US $300 per person for a 24-month program, or approximately US $12 per person per month. The original intervention cost US $189 per person, or approximately US $8 per person per month. The largest expense by far was personnel time, which accounted for approximately 72% and 64% of costs in the enhanced and original groups, respectively. Most of the increased cost in the enhanced intervention was attributable to additional personnel time for making additional monthly calls and providing SMS text message support. The prizes also contributed to higher costs in the enhanced group.

Costs and activity gains were similar throughout the first year, thus cost-effectiveness was also similar, with minutes gained in the enhanced group costing US $0.09 for each participant compared to US $0.11 for each participant in the original group (US $0.19 and US $0.16 by ActiGraph, respectively). During the second year in the program, the enhanced group continued to increase their MVPA, while gains in the original group declined. However, the original group had few costs incurred in the second year, while the enhanced group continued to deliver maintenance doses of intervention. Cost-effectiveness thus remained similar over the full 24 months, with each additional minute costing just US $0.05 in the original group and US $0.06 in the enhanced group (US $0.10 and US $0.12 by ActiGraph, respectively). ICERs showed that, beyond the cost of the original intervention, each additional minute in the enhanced intervention cost just US $0.08 (US $0.20 for ActiGraph-measured minutes). While more individuals in the enhanced group met guidelines at 24 months, the cost of meeting guidelines was also higher, costing US $812 per person compared to US $601 in the original arm.

While absolute minutes of PA varied between self-report and objective measures, as is commonly seen between these 2 measures [,], the overall pattern of results between the 2 were similar. While accelerometry is considered the gold standard, there is a large body of research showing the benefits associated with self-reported activity, which largely informed the development of national guidelines [,]. Self-reported MVPA allows for subjective interpretation of intensity, which may not align with universally applied cut points for accelerometers, particularly for participants who are overweight, obese, or inactive. Both the 7-day PAR and ActiGraph showed sustained increases in activity in the enhanced group over the 24-month study period, corroborating better maintenance of activity gains.

Comparison With Prior Work

These results suggest that paying for more intervention yields commensurate increases in activity. Given the enormous benefits of PA [,-], particularly if it is maintained over time [,], the additional cost of more intensive interventions that yield greater increase in PA over time is likely preferable for implementation sites when feasible. Although we could not find other studies that framed the incremental cost of an intervention in terms of additional people meeting PA guidelines, meeting PA guidelines is recognized as an important research metric []. One health economics study concluded that PA interventions that cost less than US $2900 over 2 years to help persons meet PA guidelines can be considered cost-effective []. Our finding that the enhanced group spent US $1837 per additional person meeting guidelines compared to the original group at 24 months is well below this threshold, and therefore, the intervention can be considered cost-effective.

This amount also seems a relatively small price to pay compared to the cost of managing chronic diseases associated with inactivity. The cost of managing diabetes, for example, was US $237 billion in the United States in 2017, with insulin alone costing approximately US $5000 per user annually [,]. An analysis of health care use in Australia found that individuals meeting activity guidelines were about one-third less likely to visit the emergency room or be hospitalized, half as likely to use outpatient services, and incurred approximately Aus $1400 (US $920) less in annual health care costs []. Multiple studies have shown that PA programs are not only low cost but ultimately cost saving, saving considerably more in health care costs than the programs cost [,]. One evaluation of a PA intervention for older community-dwelling adults found that, for those not initially meeting activity guidelines, the program saved US $143 to US $164 per participant over 6 months beyond the cost of the intervention []. Paying for PA programs can therefore be seen as an investment not only in the health and well-being of the participants but also in health care, particularly for populations at high risk. Given the higher costs and higher yields of the enhanced intervention, implementation of the enhanced version may be most appropriate with clinical populations managing chronic disease. As the original intervention still yielded substantial increases in activity, it may be more appropriate for community settings focusing on prevention and overall well-being.

Compared to the parent study, the original intervention in this study was slightly more expensive, costing US $14 per person per month at 12 months versus US $12; this was due to increased personnel costs due to wage increases []. Cost to implement the interventions will thus be largely dependent on staff salaries, which could vary broadly. Clinical sites could deliver the intervention via medical assistants or nursing staff, which would likely increase costs; conversely, using volunteers or automating components to reduce staff time could substantially lower costs.

Limitations

This study has several limitations. Cost and cost-effectiveness analyses were based on aggregated costs data, not individual cost data, thus results should be interpreted as overall estimates for delivering the intervention rather than individual effectiveness. The costs were also limited to a payer perspective and did not include health care or societal costs or quality-adjusted life years. However, these approaches allow for direct comparability with the previous study, which used the same analytic approach. Moreover, some costs, such as overhead for shared space, could only be estimated and would vary considerably based on the implementation site. We were unable to determine the effectiveness of each intervention component; therefore, it was not possible to determine the added cost-effectiveness of individual intervention components. Key differences between the conditions were the SMS text messaging and the additional calls and in-person visits. The cost of these components varied greatly, with SMS text messaging being almost free and highly scalable and calls and visits being costly and having limited scalability. Future trials using multiphase optimization strategy (MOST) designs should be used to identify the most effective components to optimize the cost-effectiveness of the intervention.

Finally, the study findings may not generalize to other populations or settings. The study population was primarily of Caribbean descent and l resided in a small geographic region of the United States. The parent study was carried out with Latinas in California who were predominantly of Mexican descent and reported being markedly less active at baseline but who showed similar increases in activity. Future research with other populations and settings would elucidate how generalizable things findings are.

The study has several strengths. Data were taken from a randomized trial with rigorous methodology, including multiple validated measures of PA. The trial focused on an underserved population considered to be at high risk and included long-term follow-up. Costs were based on current market prices, including published pay scales and benefits for staff. We were also able to directly compare costs and cost-effectiveness to those in the previous parent randomized controlled trial.

Conclusions

Findings highlight that additional intervention components, particularly those necessitating more staff time, yield higher intervention costs; however, they may also lead to greater increases in PA. In this study, these differences were most apparent after long-term follow-up. This could suggest that greater investment may be most appropriate in individuals who would benefit the most from long-term adherence to activity guidelines, such as those at high risk for, or managing, chronic diseases that respond to lifestyle changes. As staff time was by far the most costly component, PA interventions may become more cost-effective and more widely disseminated, as they rely more on broad reach technology. Future research should investigate whether relying fully on automated technology without face-to-face intervention components yields similar long-term effectiveness.

Introduction

The number of older adults is growing and expected to reach 2.1 billion worldwide by 2050 []. In Canada, approximately 18.5% of the population are ≥65 years old, representing more than 7 million Canadians [], and older Canadians constitute the fastest growing age group []. As a result, challenges have emerged with addressing the increasing needs of older adults who utilize health care services more frequently than other age groups []. eHealth, which refers to the use of information and communication technologies (ICTs) for health and encompasses a variety of mobile health (mHealth) apps [,] and internet use for health services and information delivery [,], presents an opportunity to enhance care quality, equity, efficiency, and management of health conditions among older adults [-].

In recent years, there has been a growing interest worldwide in studying technology in the context of older adults’ care [-]. With the global challenges related to the shortage of providers and the shift of care delivery outside of medical settings, older adults are expected to play a more active role in the management of their health []. In addition, as baby boomers move into the “third age” of retirement characterized by searching for new experiences and learning new things, there is an increasing need to assess their knowledge and use of information technologies [], which can support their health and well-being. This is particularly important in the context of global health care systems estimating an increase in the proportion of older adults [] and the limited resources available to support them.

The multidisciplinary life course theory (LCT) [,] emphasizes the integrated relation between a person’s choices and their socioeconomic context, alongside their capacity to make decisions within existing opportunities and constraints []. It is a multidisciplinary framework that integrates factors from various disciplines (eg, sociology, psychology) to understand human behavior. An individual’s family constitutes a “social group” that is embedded in a larger social context [], and questions related to this factor are at the core of the LCT []. The components characterizing the LCT include geographical location, social ties (eg, presence and attributes of family members and societal experiences), stages in life (eg, generational group differences), variability (eg, in gender, social class, education, wealth, family support), and personal control (eg, environmental opportunities and constraints), all of which can present circumstances that shape individuals’ perceptions and decisions []. Along the same line, the World Health Organization (WHO) emphasizes the importance of social determinants (ie, environment in which people live, work, age, and can access money and resources), as these can create avoidable differences in health across communities []. In addition, the Andersen model for health care utilization emphasizes the importance of enabling factors (eg, access to care) and need factors (eg, health problems) in shaping the utilization of health services []. Therefore, interactions with the health system, including factors related to access to resources, services utilization, health status, and perceived needs, represent contextual factors that can influence eHealth use.

Prior research on older adults’ use of technology for web-based socialization discussed the relevance of financial and knowledge barriers, as well as social factors related to family, social motivation, and appropriate environments [], and the need to investigate social and economic factors that persist as challenges to technology uptake []. Aside from a few studies that investigated social determinants in relation to telehealth use by athletic trainers [], eHealth engagement among people living with HIV [], eHealth literacy levels in older adults [], older adults’ perceptions and views on eHealth services [], and access to and preferences for patient portal use among older adults [], limited research exists in this area. In Canada, an earlier study with older adults focusing on health tracking behaviors showed that greater than 60% tracked their health manually, thus emphasizing the limited uptake of digital self-tracking in this group compared with the general population [] and the need to better understand their eHealth use patterns. Hong and Cho [], who reviewed instruments assessing eHealth behaviors, also called for national surveys adapted to technology development that may be leveraged to analyze eHealth behaviors for informing evidence-based policies.

Despite increasing scholarly interest in older adults’ use of technology, few studies have provided a comprehensive, nationally representative, and theory-driven portrait of eHealth engagement in this population. Most prior work has been limited to localized samples, focused on specific technologies (eg, patient portals, telemonitoring pilots), or relied on instruments not adapted to current technological realities. This lack of evidence impairs the ability of policymakers and health system leaders to design equitable digital health strategies that address the needs of the fastest-growing group of health care users.

Grounded in the LCT and Andersen health care utilization model, this study makes 3 important contributions. First, it provided the first national baseline of eHealth use among older Canadian adults across a broad range of applications, enabling pre- and postpandemic comparisons and international benchmarking. Second, it revealed how social determinants and health care system interactions shape digital health behaviors, offering conceptual insights into the factors that enable or constrain digital engagement in later life. Third, it highlighted equity-relevant gaps, identifying groups at heightened risk of digital exclusion and pointing to where targeted interventions are most urgently needed.

Methods

Study Design

A cross-sectional survey of Canadian older adults assessing their use of specific technologies or ICTs, eHealth, and health IT solutions was conducted online using a self-administered computer-assisted web interface (web surveys: 1500/2000, 75%) and by phone using computer-assisted telephone interviews (phone surveys: 500/2000, 25%).

Settings and Participants

A random national sample of 2000 Canadian residents from all provinces, aged 65 years and older, and who spoke English or French was selected from a proprietary online panel of more than 400,000 households owned by Léger, Canada’s largest and leading research and analytics firm. To ensure a random sample, no quotas were set initially, and the data were weighted after sampling according to gender, age, and region to maximize the representativeness of the Canadian population of older adults.

Assessments and Data Sources

In line with the conceptual framework shown in , the survey instrument included 3 main sections: (1) sociodemographic characteristics, living environment, and health system interaction factors; (2) eHealth and ICT use; and (3) fall detection technology (FDT) and telemonitoring technology (TM) use.

‎

Grounded in the principles of the LCT, sociodemographic variables were measured with standardized indicators including gender, age, marital status, income, education, language, and employment. Questions assessing the living environment included the region (representing provinces), community (rural/suburb/urban/metropolitan), place of residence (home/retirement home/long-term care), and living arrangement (alone/with family/with spouse or partner) to give a comprehensive overview of the older adult social context.

To measure interactions with the health care system, we used the following indicators: private insurance (yes/no), family physician (yes/no), home care (yes/no), willingness to pay for quicker access to services (yes/no), hospitalizations and emergency room visits in the past 6 months (yes/no), and perception of own health (categorical). Satisfaction with the health care system, with access to health care services, and with the health care services received were also included in the survey (categorical).

The outcomes of interest in this study were divided into 3 categories that represent eHealth use: (1) digital health engagement, including the extent of internet use for actively searching and accessing health-related information and services online (eg, searching for online information about a health condition, accessing laboratory test results); (2) digital health communication, including the extent of willingness or interest in using digital means to communicate about health services (eg, use of email to discuss a health condition with a physician, obtain information about trusted websites); and (3) mobile health technology use, including the extent of use of mobile apps, FDTs, and TMs. The questions assessing FDT and TM use were binary (yes/no). The frequencies of use of mobile apps for health and internet for accessing information, resources, and communicating with providers were measured on a 5-point scale (1=Never to 5=Always). Interest or willingness to exchange online information was also assessed on a 5-point scale (1=Not at all to 5=Totally).

The survey instrument was pretested with 47 respondents by phone and on the web. A change was subsequently made to the skip pattern for one of the questions. Data collection was completed over a 3-week period in 2018. Although the data are not very recent, they provide a robust baseline for pre- and post–COVID-19 pandemic comparisons, especially in the absence of data on eHealth use among older adults.

Data Analysis

Descriptive data analysis was conducted to gain an understanding of the profile of older adults and their technology-related behaviors. Bivariate nonparametric tests were used for analyses of associations between continuous dependent variables (eg, search online for information) and categorical independent variables (eg, age, education). Specifically, nonparametric 2-independent samples analyses (Mann-Whitney) were conducted for independent variables with only 2 categories, and nonparametric K-independent samples analyses (Kruskal-Wallis) were conducted for independent variables with more than 2 categories. χ² and Fischer’s exact tests for categorical variables were used to examine the associations between the binary dependent variables related to mobile technology use and the determinants (eg, sociodemographic characteristics, living environment) and interactions with the health system variables. Multivariate regression analysis for eHealth scale questions and binary logistic regression analysis for binary mHealth technology questions (TM and FDT) were performed to examine the significant relationships between eHealth use and the social and health system interaction variables that showed significant associations at the bivariate analysis level. To assess potential multicollinearity among the independent variables for the multivariate analysis, we calculated the variance inflation factor (VIF) for each predictor. All independent variables, with the exception of marital status (ie, “married”; VIF~10), had a VIF lower than the general threshold of 10 []. Therefore, we retained all theoretically relevant variables to maintain the breadth of the models, which was the main aim with our modeling approach.

Ethical Considerations

Ethics approval was granted by the University of Ottawa Ethics Research Board, and all data were anonymized. Eligible respondents were provided with an information letter that explained the scope and purpose of the project. It specified that their participation was voluntary and completing the questionnaire indicated their consent for participating in this study. Léger panelists are rewarded for their participation over time using a series of financial incentives that can be accumulated and cashed out or donated for a charitable cause. Participants in this survey who completed this survey were rewarded 1200 Léger Opinion points (CAD $20=20,000 points; a currency exchange rate of CAD $1=US $0.71 is applicable) that can be redeemed through their preferred reward method.

Results

Sample Characteristics

reveals the variations in social determinants and living conditions. The majority of respondents were 65 years to 69 years old (663/2000, 33.2%), were women (1092/2000, 54.6%), were married (1201/2000, 60.1%), had a college or university degree (1243/2000, 62.2%), spoke English at home (1496/2000, 74.8%), earned less than CAD $75,000 (1184/2000, 59.2%), and lived in Ontario (760/2000, 38%) or Quebec (505/2000, 25%), which are the most populous and largest health care jurisdictions in the country. In addition, 68.5% (1369/2000) lived in metropolitan cities or suburbs, and 96.1% (1922/2000) lived in their own homes or apartments; one-third (665/2000, 33.3%) lived alone.

Health Care System Interaction

Of the repondents, who normally have access to a national health insurance, 54.9% (1098/2000) reported having private insurance (). The majority had a regular family physician (1896/2000, 94.8%), did not receive home care services (1917/2000, 95.9%), and reported good to excellent health (1661/2000, 83.1%). In addition, 8.2% (164/2000) and 13.9% (278/2000) had hospitalizations and emergency visits, respectively, in the past 6 months, and one-third (647/2000, 32.4%) were willing to pay out of pocket for quicker access to services. Overall, 61.8% (1237/2000) reported being diagnosed with one or more chronic conditions, and 13.7% (275/2000) indicated having fallen in the past 6 months (falls: median 1, IQR).

Overall, the satisfaction of respondents with the health care system was high (very satisfied: 366/2000, 18.3%; satisfied: 1086/2000, 54.3%). In addition, 79.1% (1583/2000) indicated good to high satisfaction with access to health care services, and 84.4% (1689/2000) were satisfied to very satisfied with the health care services received over the past 2 years.

ICT and eHealth

The vast majority of surveyed older adults owned a computer (1703/2000, 85.2%), 57.7% (1153/2000) reported having a tablet or iPad, and 53.9% (1077/2000) reported having a smartphone (). Fewer owned wearables or mobile devices (238/2000, 11.9%). In addition, 90.3% (1654/2000) reported using email, 50.1% (917/2000) used phone text messaging (eg, WhatsApp, Messenger), and 53.6% (983/2000) used Facebook. Except for a small percent of respondents (238/2000, 11.9%), participants in our study confirmed having used the internet over the past 6 months, mostly daily (1472/2000, 83.6%) or a few times a week (180/2000, 10.2%).

Despite frequent internet and email use and high prevalence of ICTs, the use of internet to connect with a health care professional, access test results or patient portal, or book a medical appointment was limited, although moderate use was reported for using the internet to search for online information about health conditions (median 3, IQR 2-3 on the 5-point Likert scale; ). The respondents expressed low willingness to use email to exchange information about their health with their health care professionals (median 2, IQR 1-3 on the 5-point Likert scale) and a moderate interest in obtaining information about trusted websites relevant to their health condition and accessing their medical records (median 3, IQR 1-5 on the 5-point Likert scale).

The prevalence of use of TMs and FDTs was low (189/2000, 9.4% and 84/2000, 4.2%, respectively), despite familiarity with these technologies. Among the respondents who had previously or were currently using TMs and FDTs, 81% (153/189) and 86% (72/84), respectively, indicated their willingness to use these technologies again in the future, which indicates satisfaction with these technologies.

Social Determinants, Health Care System Interaction, and eHealth Use

Bivariate analyses showed significant associations (P<.05) between sociodemographic and living environment characteristics and the majority of questions assessing eHealth use (). For example, being a man, in the younger age group (ie, 65‐69 years), and holding a graduate university degree were significantly associated with a higher willingness or interest in using the internet to access medical records, exchange medical information with family physicians or health care providers, and obtain information on trusted websites to consult about one’s condition (all P<.001). On the other hand, being a woman, being separated or divorced, and having a lower income were significantly associated with higher frequency of internet use to search for online information about health problems (all P<.001) and the tendency to self-diagnose (P<.001, P=.02, and P=.03, respectively). Living in metropolitan areas and in one’s own home were significantly associated with higher frequency of using mobile apps for health (P=.03) and TM use (P=.04). Being married or having a partner was also significantly associated with using mobile apps for health (P=.01) and FDT use (P=.01).

When examining the relationship between interactions with the health care system and eHealth use (), having private insurance and a willingness to pay out of pocket for quicker access to health care services were significantly associated with most outcome variables (eg, most P≤.001). Respondents who indicated not receiving home care services and no hospitalizations nor emergency visits reported more FDT use (P<.001, P=.002, and P=.002, respectively). Those having an excellent perception of health reported more use of mobile apps for health, higher frequency of internet use to participate in discussion forums about their health, and more willingness or interest in using email to exchange medical information with their physician about their condition.

Multivariate Analysis

Multivariate analyses examined the relationship between eHealth use and the sociodemographic characteristics and interactions with the health care system variables that were significant at the bivariate analysis level while controlling for other variables. presents which relationships were significant (P<.05) in the multivariate analyses; the detailed results (standardized coefficients, odds ratios [ORs], confidence intervals, and P values) are presented in .

eHealth use (ie, digital health engagement and digital health communication) was significantly associated with several sociodemographic variables (). Women reported more internet use to search for online information about health problems (β=0.148, P<.001) and self-diagnosis (β=0.079, P=.001), whereas older age (≥85 years) was consistently associated with lower frequency of eHealth use (internet use for self-diagnosing, interest in obtaining information about trusted websites for their health condition, accessing online medical records, and use of mobile apps for health). Interestingly, respondents with a lower income indicated a higher frequency of searching for online information about health conditions or problems (β=–0.122 for those earning between CAD $50,000 and CAD $75,000 compared with those earning less than CAD $25,000; P=.007), and English-speaking respondents (compared with their French-speaking counterparts) had a higher frequency of self-diagnosing (β=0.098, P<.001) and participating in online forums to discuss aspects related to their health (β=0.051, P=.04).

Variation in eHealth use was significantly associated with the region and community of residence. Compared with rural areas, living in the suburbs or metropolitan areas was consistently associated with a higher frequency of using mobile apps for health (β=0.135, P<.001 for metropolitan areas) and internet use for looking for information about health conditions (β=0.130, P<.001 for suburban areas), self-diagnosing (β=0.114, P<.001 for suburban areas), accessing laboratory results (suburban areas: β=0.092, P=.006; metropolitan areas: β=0.077, P=.03), accessing patient portals (suburban areas: β=0.088, P=.01; metropolitan areas: β=0.077, P=.35), and booking appointments online (β=0.105, P=.005 for metropolitan areas). Residing in retirement homes as opposed to one’s own home was significantly associated with more FDT use (OR=0.366, 95% CI 0.145‐0.923).

There were considerable differences across provinces in Canada, which may be attributed to systemic variation in availability of digital health services leading to variable access and eHealth use. When compared with persons from central Canada (ie, Prairies), Canadians residing in British Columbia reported a higher frequency of using the internet to ask health care professionals about their health (β=0.062, P=.04) and mobile apps for health (β=0.064, P=.04) and more willingness or interest in using email to exchange information about their health condition with a physician (β=0.087, P=.002) and access their online medical records (β=0.083, P=.002). Residents of Ontario also reported a higher frequency of using the internet to search for health information about their conditions (β=0.075, P=.03) and access laboratory results (β=0.297, P<.001) and patient portals (β=0.132, P<.001), whereas respondents from the Maritimes provinces were more interested or willing to use the internet to email their physicians about their health condition (β=0.093, P<.001), obtain information on trusted websites to consult about their conditions (β=0.062, P=.01), and access their online medical records (β=0.089, P<.001). Current use of the internet to access patient portals (β=0.082, P=.01) and interest in accessing online medical records (β=0.120, P=.005) were also higher for older adults in Quebec compared with residents of central Canada.

The enabling and need factors related to the interactions with the health care system investigated in this study revealed a pattern of association with eHealth use. Among the enabling factors, willingness to pay out of pocket for quicker access to health care services and having private insurance were consistently and significantly related to a higher frequency of eHealth use. Specifically, willingness to pay out of pocket was significantly associated with higher frequency of eHealth use for all the measures, with the exception of mobile app and FDT use (). Respondents who did not have private insurance reported lower frequency for searching for information about their health problem or condition (β=–0.054, P=.03) and online access to laboratory results (β=–0.055, P=.02) and patient portals (β=–0.063, P=.009). With regard to the need variables, older adults who did not have emergency visits in the past 6 months (ie, less needs) reported a significantly lower frequency of searching for information about their health problem or condition (β=–0.086, P<.001) and accessing patient portals (β=–0.058, P=.02) but more FDT use (OR=2.16, 95% CI 1.228‐3.800). However, those who did not receive home care services indicated a higher frequency of searching for information about their health problem or condition (β=0.052, P=.03) and more FDT use (OR=3.427, 95% CI 1.550‐7.596) compared with those receiving home services. Last, a high perceived health status (as excellent) was significantly associated with more frequent mobile app use as opposed to a fair (β=–0.073, P=.02) or good (β=–0.074, P=.049) perception of health.

Discussion

Principal Findings

This national survey provides the first comprehensive portrait of eHealth use among older Canadians, offering a valuable baseline for ongoing monitoring and international comparisons. Although the analyses incorporated a broad range of social and health system determinants, our goal was to capture the multidimensional nature of digital engagement and application use in later life, which sets the stage for future specific and targeted investigations. To enhance clarity, the Discussion section highlights the most important findings and their implications.

First, although the vast majority of older adults owned digital devices and reported frequent internet use, their actual engagement with eHealth tools remained limited. Use of TMs, FDTs, patient portals, and online appointment systems was low across the sample. This disconnect suggests that access alone is insufficient; awareness, perceived usefulness, and provider support are critical enablers. Importantly, those who had used TMs and FDTs expressed high willingness to use them again, underscoring the importance of initial exposure and positive experiences.

Second, consistent with prior literature, eHealth use was stratified by sociodemographic advantage. Younger age, higher income, and residence in metropolitan areas were associated with greater engagement, while older age, lower income, rural residence, and institutional living environments were linked to reduced use. These patterns reveal persistent inequities within the older adult population, even in a context of high technological readiness. Targeted efforts are needed to ensure that those with the greatest health needs and fewest resources are not left behind.

Third, interactions with the health care system emerged as powerful predictors of eHealth use. Having private insurance, willingness to pay for quicker access, and better perceived health were all associated with higher digital engagement. Conversely, those who had received home care services or experienced recent emergency visits were less likely to use digital tools. This pattern indicates that eHealth may currently serve those with fewer immediate health needs and greater resources, rather than those most in need of coordinated care. Integrating digital health into routine pathways, rather than leaving it as an “optional extra,” is essential for equity.

Fourth, women reported greater use of the internet for health information and self-diagnosis. Although our data do not allow us to assess outcomes such as health anxiety, prior research has suggested that extensive online searching can sometimes heighten distress (ie, “cyberchondria”) [,]. This represents a possible area for future investigation.

Several additional associations were observed, including the roles of education, language, and provincial differences, that further illuminate the diversity of older adults’ digital health engagement, which was quite variable across the country. For example, residents of Ontario and British Columbia reported higher use of patient portals and access to laboratory results, while those in the Maritimes expressed greater willingness to communicate with providers online. These contextual nuances reinforce the importance of tailoring digital health strategies to local and cultural environments.

Comparison With Studies in Other Jurisdictions

Our findings contribute to a growing body of international literature on digital health use among older adults and offer several contrasts with studies conducted in the United States and Europe.

First, older Canadian adults in our sample reported higher levels of interest and perceived usefulness of digital health tools across age groups, including those older than 75 years. This contrasts with US and European studies that consistently report lower adoption rates among the oldest and most socioeconomically disadvantaged groups [-]. Second, although previous studies often focus on access and skill gaps, the so-called first- and second-level digital divides, our findings highlight a third dimension: motivational readiness. Many respondents expressed a willingness to use digital tools despite limited experience, particularly when they felt supported by the health care system. This dimension was less frequently emphasized in prior large-scale survey studies []. Third, the Canadian context appears to moderate some of the demographic divides found elsewhere. For example, racial and ethnic disparities in digital health use reported in the United States [] were not observed in our sample, potentially reflecting Canada’s publicly funded and more equity-oriented health care system. Finally, our data indicate relatively high willingness to reuse certain digital health technologies, particularly TMs and various mobile apps. This is an encouraging finding given the infrastructural and digital literacy barriers reported in many European contexts [,] and suggests that, when appropriately introduced and supported, older adults can develop positive experiences with digital health tools, even in rural areas.

Taken together, our findings suggest that older Canadian adults are not uniformly resistant to digital health but rather face a complex set of motivational, attitudinal, and structural barriers. Understanding these factors in light of international comparisons can inform more targeted, inclusive, and equity-driven eHealth policies.

Study Implications and Avenues for Future Research

The findings of this national survey of older Canadians have several implications for policy, practice, and future research on digital health equity. First, the results underscore the need to move beyond binary notions of access (eg, having internet or not) to understand the multidimensional nature of digital engagement. Although structural factors such as age, education, and income remain important, our findings reveal that attitudinal variables such as perceived usefulness, trust in digital tools, and self-efficacy are equally, if not more, influential in predicting eHealth adoption. This suggests that interventions should not only address material barriers but also focus on building digital confidence and relevance in health contexts.

Second, the relatively high levels of expressed interest in eHealth technologies among Canadian older adults, even those with limited prior experience, challenge persistent narratives of older adults as digitally disengaged or resistant. This observation contrasts with trends reported in European contexts [,], where digital disengagement remains widespread, particularly among the oldest-old, women, and rural residents. Similarly, in the United States, studies by James et al [] and Schuster et al [] identified persistent digital divides by race and mental health status. In contrast, our findings point to a more complex and optimistic outlook in Canada, where universal health care, relatively equitable access to health services, and targeted digital health initiatives may be shaping more inclusive digital health trajectories.

Third, the Canadian context provides a unique lens to understand how publicly funded health care systems may buffer against some of the exclusionary forces observed in more market-based systems. For instance, unlike in the United States, where the use of digital health tools is often mediated by private insurance coverage or provider-specific platforms, Canadian respondents interact with a publicly funded system that, in principle, offers more universal access to tools such as e-prescriptions and teleconsultation portals. However, our findings indicate that actual use of digital services like these remains limited, suggesting that accessibility alone does not ensure engagement and that awareness, support, and perceived usefulness remain critical enablers.

Based on these findings, several avenues for future research emerge. Longitudinal studies are needed to assess how digital engagement evolves over time among older adults, especially as younger cohorts with greater baseline digital skills age into retirement. Moreover, future research should examine how relational dimensions, such as the support of health care professionals, caregivers, and peers, mediate eHealth use among older adults. Comparative studies across health systems and sociopolitical contexts would further elucidate the interplay of systemic and individual-level factors in shaping digital engagement. Building on these findings, targeted investigation of specific eHealth applications using mixed methods and qualitative approaches would provide a more in-depth understanding of the interaction of factors influencing their use. Finally, there is a need to develop and evaluate intervention strategies that go beyond digital literacy training. These should include motivational and psychosocial components, co-designed with older adults, to enhance perceived value and usability of digital tools in real-life care scenarios. Integrating such approaches into existing health care pathways can ensure that digital transformation in health systems is truly inclusive, responsive, and sustainable for an aging society.

Study Limitations

It is important to note some limitations associated with this research. The cross-sectional nature of this study precludes a thorough assessment of causal relationships between the social determinants and variables related to the interactions with the health care system in relation to eHealth use. For example, the odds of using FDTs were considerably higher among older adults who did not receive home care services nor were hospitalized in the past 6 months. However, it was not possible to determine whether FDT use precluded the need for home care and prevented hospitalizations or whether FDT was used due to the absence of home services and better health.

The respondents’ profile points to a relatively high level of education (62% had a college degree or higher) with the majority residing in their own homes and not alone, living in suburbs or metropolitan areas, having a regular family physician, and having private health insurance (ie, generally good access to a broad range of health care services despite limited use of home care services). Thus, we may expect a lower level of eHealth penetration among the broader older adult population, which further underscores the current suboptimal benefits that older adults are gaining from these technologies.

Since the data were collected from a single country, the generalizability of the results is limited unless the survey is replicated in other contexts. The online nature of the survey, although complemented with phone surveys, may still have excluded potential respondents who did not have access to the internet or phone calls. In addition, the closed-ended questions included in the survey did not allow us to fully uncover the reasons behind some of the association patterns that were observed.

When assessing multicollinearity, marital status (ie, “married”) had a VIF close to 10, indicating potential multicollinearity with other predictors in the models. This is expected, as marital status is closely associated with other sociodemographic variables like cohabitation. Although this multicollinearity may affect the precision of the estimated effect of marital status (ie, compresses the ß coefficient and makes it more likely to find false negatives or more conservative results), it does not bias the model overall.

Last, we must stress that there are persistent challenges with collecting comprehensive data from older adult populations [,,]. Although the data from this study are not very recent, they provide a baseline for future studies assessing changes in eHealth use among older adults post-COVID-19 pandemic. In addition, since the eHealth construct’s evolution is gradual and eHealth use among Canadian citizens in general is reported to be changing slowly [], the findings continue to be relevant and present a foundation for longitudinal comparisons.

Conclusion

This study offers the first comprehensive national assessment of eHealth use among older Canadian adults and provides a valuable baseline for ongoing monitoring and international benchmarking, including pre- and postpandemic comparisons. Our findings highlight the importance of accounting for social determinants and interactions with the health care system when investigating eHealth use in this population.

Our findings are also relevant beyond Canada. They demonstrate how universal health care systems mitigate, but do not eliminate, digital divides, offering lessons for other jurisdictions seeking to advance inclusive digital health strategies. The study reveals a new dimension of the digital divide, namely motivational readiness, which complements traditional access- and skill-based divides. Recognizing this attitudinal and relational component shifts the focus of interventions from infrastructure alone to trust-building, perceived usefulness, and provider support, thereby broadening the policy tool kit for promoting digital equity in aging societies.

Introduction

Atherosclerotic cardiovascular disease (ASCVD) remains one of the top global health concerns []. It is estimated that 80% of ASCVD cases are preventable []. Strategies to identify and target people at higher risk are therefore crucial to reducing the global disease burden, especially in resource-constrained settings with limited access to medical specialists and advanced equipment [,].

Primary care doctors play an important role in preventing ASCVD by assessing and managing cardiovascular risk factors. Risk assessment is centered on routine screening of asymptomatic adults 40‐75 years of age to determine the risk of heart disease or stroke in the next 10 years, allowing for more focused primary preventive interventions. The intensity of interventions must align with the patient’s individual risk, optimizing anticipated benefits while minimizing potential harm from overtreatment and allocating limited resources appropriately []. Doctors often use clinical decision support (CDS) tools to calculate ASCVD risk and guide shared decision-making for preventive management [,]. In resource-constrained settings however, medical records are predominantly maintained on paper, and doctors use CDS on mobile devices [] (AP Susanto, PhD, et al; unpublished data, August 2024). This situation warrants an automated process to enter patient data into the CDS, as manual data entry is prone to errors, especially under high patient loads (AP Susanto, PhD, et al; unpublished data, August 2024).

Recently, CDSs are being embedded with artificial intelligence (AI) [], specifically machine learning (ML) models that can accommodate patient-level precision beyond traditional risk factors. These AI-based CDSs provide new opportunities to improve risk assessment in poorly served patient subgroups and consider nontraditional risk factors, such as psychosocial factors, or inflammation [,]. While ML models have demonstrated superior performance compared to statistics-based ASCVD risk calculators [], their impact on enhancing clinical decisions, which is an essential precursor for realizing benefits in care delivery and patient outcomes, remains unknown []. In a recent review of AI tools in routine care, only 7 studies examined clinical decision-making, none of the 86 randomized controlled trials published in the last 6 years focused on risk assessment for cardiovascular disease (CVD), and only 4 were conducted in resource-constrained settings [].

We sought to evaluate whether AI-based CDS can improve clinical decisions for the primary prevention of ASCVD in Indonesia, where CVD accounts for the highest health care expenditure []. The high prevalence of CVD risk in Indonesia (29.2%) underscores the need to prioritize preventive services []. Accurate risk assessment is fundamental to preventive care, yet it remains a persistent challenge, especially among high-risk rural residents who are less likely to receive such care. Consequently, many individuals at elevated risk are not identified, whereas others without risk are unnecessarily subjected to preventive interventions. A randomized controlled study was conducted to assess the effect of AI-based CDS on 10-year ASCVD risk assessment and management, using clinical vignettes with a conceptual prototype of a CDS. Clinical vignettes or patient cases offer greater study control to quantify the effects of AI-based CDS on decision-making before costly and potentially disruptive clinical deployment, allowing for evaluation without putting real patients or doctors at risk [,]. The patient cases were based on our previous work, which examined the sociotechnical context of CVD risk assessment in resource-constrained settings with limited access to medical specialists and advanced equipment (AP Susanto, PhD, et al; unpublished data, August 2024). Acute coronary syndrome caused by atherosclerosis was confirmed as the most common CVD challenge, requiring primary prevention to alleviate the disease burden. In this study, we compared AI-based CDS against the current approach to risk assessment with either no CDS or automated CDS, whereby patient data were automatically populated into the CDS. The technology acceptance of CDS was also examined.

Methods

Setting and Participants

In total, 102 doctors participated in the online study between November and December 2023. Participants were recruited by publishing a call for volunteers via the Indonesian Heart Association at 5 training centers located in different geographical regions. Doctors who were working in general practice or enrolled in a cardiology training program were eligible to participate. The study invitation was also distributed to potential participants via email and WhatsApp (Meta Platforms).

Design

We performed a 3-way, within-subject randomized controlled study to evaluate the effect of CDS type: standard care with no CDS, automated CDS, and AI-based CDS (see and the Intervention section for details). To simulate variation in the clinical context, each CDS type was provided with 3 patient cases with different case complexity levels: low, high, and high with risk-enhancing factors (REFs) based on the literature [,]. This design provided 9 patient cases related to primary prevention in an outpatient setting for adults aged 40-75 years without a history of ASCVD (AP Susanto, PhD, et al; unpublished data, August 2024). Low-complexity cases only contained information about the traditional risk factors, including age, sex, race, blood pressure, cholesterol level, history of diabetes, smoking status, and previous medications including antihypertensives, statin, and aspirin. High-complexity cases included the presence (3 cases of high complexity with REFs) or absence (3 high-complexity cases) of REFs such as family history of premature heart attack and obesity, on top of the traditional risk factors.

‎

The study was deployed using Gorilla.sc, a platform for conducting online studies, which was customized to display patient cases and CDS interventions []. The allocation of cases to the different CDS types and the presentation order were randomized so that each case would appear evenly across the 3 CDS types. All randomizations were set equal (1:1) and performed by the Gorilla software (v2; Cauldron Science).

Intervention

The CDS type provided to participants was manipulated across 3 conditions as follows:

1. No decision support, that is, the current approach to risk assessment served as the control condition. Only the information present in the case was shown (see ).
2. Automated CDS: This was designed to augment the current approach to risk assessment whereby patient data were automatically populated, and the CDS output was displayed to participants, streamlining screening for doctors with high patient load. The automated CDS presented the 10-year ASCVD risk level, risk score (in %), and treatment advice in English. It was based on the publicly available ASCVD Risk Estimator Plus calculator from the American College of Cardiology [], which is widely accepted in Indonesia. Risk is calculated using Pooled Cohort Equations, a conventional statistical method based on traditional risk factors. Although the display of the original calculator was not altered, the logo was concealed (see ).
3. AI-based CDS: This was designed to represent contemporary risk calculators based on traditional ML models (eg, random forest, neural networks, logistic regression) that account for additional risk factors on top of the traditional risk factors and have been shown to perform better in risk discrimination compared to traditional risk estimators []. Here, a conceptual prototype of an AI-based CDS was constructed similar to previous studies examining doctors’ perceptions about ML-based risk calculators []. The AI-based CDS presented the 10-year ASCVD risk level and treatment in English but did not provide a risk score (in %). The CDS output, that is, the risk assessment and treatment advice for each case, was constructed by a general practitioner (GP; APS), then scored independently by 2 consultant cardiologists (DAJ and AS), and reviewed by a third consultant cardiologist (BW) based on their clinical experience in managing patients with CVD in Indonesia. All disagreements were resolved by consensus in a workshop (see and ).

‎

Blinding was implemented to minimize placebo effects. The allocation of patient cases to CDS type was according to 9 pregenerated sequences, created so that cases were allocated evenly across each of the 3 conditions. Participants were allocated to sequences at the time of enrollment using balanced randomization, with sequences presented in random order. All conditions were displayed using an identical user interface. While participants were able to distinguish between the control (no decision support) and intervention conditions, they were blinded to the intervention conditions. The automated CDS and AI-based CDS were generically coded as “Risk Calculator 1” and “Risk Calculator 2,” respectively, with an identical presentation (). We did not specify which tool harnessed AI to minimize placebo effects that could influence doctors’ performance []. In all conditions, participants could access 2 clinical guidelines on ASCVD prevention [,], commonly referred to by Indonesian doctors [].

Patient Cases

The patient cases were based on the clinical guidelines for the primary prevention of CVD [,], covering typical presentations of asymptomatic adults 40‐75 years of age in an outpatient setting. Participants were presented with 9 patient cases and asked to assess the 10-year risk of ASCVD and provide the most appropriate patient management. As shown in , each case included a brief patient history, physical examination, and simple laboratory results commonly available in resource-constrained clinical settings (). The study setup and the CDS interface were designed to emulate access on a mobile device, which is a likely real-world implementation of such tools in resource-constrained settings [].

Cases and questions were presented in Bahasa Indonesia to ensure clarity. The cases were developed by a GP (APS) with advice from a senior cardiologist consultant (BW). Risk factors, risk assessment, and patient management in each case were coded by APS and reviewed by BW; disagreements were resolved by consensus. The gold standard for risk assessment was determined based on the clinical guidelines, and cases with REFs were reclassified based on clinical relevance. Cases and gold standard responses for correct risk assessment and patient management were independently reviewed by an expert panel of 2 senior cardiology consultants (DAJ and AS) to ensure clinical relevance. Disagreements were resolved by consensus via a finalization workshop.

Procedure

Participants self-enrolled in the online study. The study URL was provided in the invitation, and participants were asked to access the study from a laptop or desktop computer with an internet connection at their convenience. After informed consent, participants provided demographic information and watched an instructional video (2 min) explaining the study task and CDS interventions, orienting participants with the study user interface.

Participants were also shown how to view clinical guidelines that were accessible online. At the end of the video, participants had the opportunity to explore the user interface and CDS interventions with a demonstration case and could repeat the demonstration case once.

Participants were randomly allocated a pregenerated sequence of 9 cases. Cases that were allocated to the automated CDS (blinded as “Risk Calculator 1”) and AI-based CDS (“Risk Calculator 2”) conditions presented the risk level and patient management. Control cases displayed “No Decision Support” where CDS would not be displayed in the mobile device interface.

Participants were instructed to complete all 9 cases to the best of their clinical judgment in a single, uninterrupted session, and no time limits were imposed. After completing the cases, participants responded to a poststudy survey based on the technology acceptance model (TAM) []. They were then provided with a debriefing note and thanked for their participation. The procedure was pilot-tested by 2 doctors and refined based on their feedback.

Outcome Measures

To assess the effects of AI-based CDS on risk assessment and patient management, the following outcome measures were examined.

Risk Assessment

The response was multiple choice and automatically scored as correct or incorrect against the validated gold standard response for the cases. Given the case, participants determined the 10-year ASCVD risk assessment, defined as a chance of having major CVD events (eg, heart disease or stroke) in the next 10 years. There were 4 levels of risk: low risk (<5%), borderline risk (≥5% to <7.5%), intermediate risk (≥7.5% to <20%), and high risk (≥20%). We counted the number of cases correctly assessed.

Patient Management

As with the risk assessment, the response for patient management was multiple choice and automatically scored as correct or incorrect against the validated gold standard response for the case. Given the case and risk classification, there were 4 measures of patient management for primary prevention: whether the doctor discussed the risk level or score and prescribed aspirin, statins, antihypertensives, and referred patients for advanced examinations. We counted the number of cases correctly managed as follows.

1. Aspirin prescription: Participants chose between not prescribing or prescribing 80‐100 mg of aspirin.
2. Statin prescription: Participants choose 1 of 4 options: (1) no statin therapy, (2) low-intensity statin (eg, simvastatin 10‐20 mg), (3) moderate statin intensity (eg, simvastatin 20‐40 mg or atorvastatin 10‐20 mg), or (4) high-intensity statin (eg, atorvastatin 40‐80 mg or rosuvastatin 20‐40 mg).
3. Antihypertensive prescription: Participants chose between discussing and prescribing antihypertensive medications or not.
4. Advanced examination referral: Participants chose between referring and not referring patients to a facility located 1‐3 hours away for advanced examinations such as treadmill stress test [], computed tomography calcium score [], or computed tomography coronary angiography [].

Decision-Making Time

Decision-making time was automatically measured from when the case was first presented until the responses for risk assessment and diagnosis were submitted.

Technology Acceptance

Participant perceptions about the potential utility of AI for ASCVD risk assessment in resource-constrained settings were examined using mixed methods []. In the quantitative component, we examined perceptions about the automated CDS and AI-based CDS using the TAM, including the usefulness (2 items), ease of use (1), and behavioral intention of CDS usage (2) using a 7-point Likert scale ranging from (“strongly disagree”) to 7 (“strongly agree”). Participants’ free-text responses to “comments about CDS” were coded into different factors of the TAM by APS and reviewed by FM.

We also measured effects of case complexity, diagnosis, choice of specific antihypertensive regimen, and justification for referrals; these analyses will be reported separately.

Statistical Analysis

We estimated that 61 participants would be required to detect a 39% or greater difference (2-tailed) in decisions with and without CDS with 90% power and at P<.05 []. All measures for risk assessment and patient management were non-normally distributed, and therefore, the effects of CDS were tested by using the Friedman test to compare multiple repeated measures of cases correctly assessed and managed across CDS types. Subsequent CDS pairwise comparisons with Bonferroni-adjusted P values and effect sizes were performed between the CDS type pair.

Decision-making time was analyzed by a multilevel model with outliers removed (see ) [,]. The model included a random intercept for each participant, controlling the nested nature of data. Predictors assessed for inclusion in the model were CDS type and case complexity. A stepwise backward elimination method was used for predictor selection, where all predictors and interactions were entered into the model. Models were estimated using maximum likelihood, and fit was evaluated. Predictors that significantly improved model fit were retained.

For technology acceptance, the median of all participant scores was calculated for each item. To assess the internal consistency of the TAM factors in the questionnaire, we calculated the mean of the items for each factor and Cronbach α [].

All statistical analyses were undertaken using the SPSS software (v27; IBM Corp). Reporting was guided by the health care simulation research extensions of the CONSORT (Consolidated Standards of Reporting Trials) and STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statements (see ) [].

Ethical Considerations

This study was approved by the Macquarie University Human Research Ethics Committee (reference: 520221032837485; project: 10328) and Faculty of Medicine Universitas Indonesia Ethics Committee (KET-586/UN2.F1/ETIK/PPM.00.02/2022). Participant consent was obtained via an online participant information sheet and consent form in accordance with approved protocols. Participants were offered an IDR 500,000 (US $32) e-commerce voucher for their contribution. No personally identifying information was used for analysis and reporting. Patients presented in the cases were hypothetical, biographical information was generated solely for the purposes of this study, and patient photos were produced by generative AI (Dall-E by OpenAI in October 2023). This study was not prospectively registered, as it was conducted using a conceptual prototype with clinical vignettes rather than real patients. It was retrospectively registered with the Australian New Zealand Clinical Trials Registry (ACTRN12625001168448 []).

Results

Demographic Characteristics

In total, 102 Indonesian doctors participated in the online study (; ). The participants were distributed across all 7 geographical units of the Indonesian archipelago (). Most (n=57, 56%) identified as male and were aged 26‐35 years (n=85, 83%), with a median of 6 (IQR 4.75) years of clinical experience. Of these, 63 (62%) were enrolled in a 4-year graduate cardiology residency training program, which requires 1‐5 years of experience as a GP following graduation from undergraduate medical school and is usually undertaken in a remote area.

‎

‎

AI-Based CDS Improved Risk Assessment

We examined the effect of CDS type on risk assessment by comparing cases correctly assessed across the 3 study conditions: standard care with no CDS (control), automated CDS, and AI-based CDS (). The analysis revealed a statistically significant difference in risk assessment by CDS type (: P<.001). In pairwise comparisons, a significantly higher number of cases were correctly assessed when doctors were assisted by the AI-based CDS compared to the control (z=−5.602, n=102, adjusted P<.001, medium effect size r=0.39) and the automated CDS (z=−4.901, n=102, adjusted P<.001, medium effect size r=0.34). However, the improvement in risk assessment with the automated CDS compared to the control was nonsignificant (z=−0.700, n=102, adjusted P>.99).

Effects of AI-Based CDS on Patient Management Were Mixed

We observed a mixed effect of CDS type on the 4 patient management measures, including the prescription of aspirin, statin, antihypertensive medications, and referral for advanced examinations ( and ). While CDS type had a statistically significant effect on the prescription of statins (P<.001) and referrals (P=.03), its effect on the prescription of aspirin and antihypertensives was nonsignificant.

In pairwise comparisons, a significantly higher number of cases were prescribed with the appropriate statin intensity when doctors were assisted by the AI-based CDS (z=−4.936, adjusted P<.001, medium effect size r=0.35) and the automated CDS (z=−3.991, adjusted P<.001, small effect size r=0.28). For referrals, no differences were revealed in the pairwise comparison. While fewer cases were correctly referred with AI-based CDS compared to the control (z=0.230, adjusted P=.30) and automated CDS (z=2.135, adjusted P=.30), these differences were nonsignificant.

AI-Based CDS Decreased Decision-Making Time

Cases assisted by the AI-based CDS (P=.02) and automated CDS (P=.006) took less time than the control (see for estimated marginal means). There was no significant difference in the decision-making time between cases assisted by the automated CDS and AI-based CDS.

Perceptions About CDS

We measured doctors’ perceptions about automated CDS and AI-based CDS using the TAM. Overall, doctors agreed with statements measuring the perceived ease of use, usefulness, and behavioral intention to use CDS, with a median score of 6 for all items (ie, “agree”). The value for Cronbach α for the survey was ≥.7, showing internal consistency within the factors. Of the 102 respondents, 81 (79%) agreed or strongly agreed to using recommendations from AI-based CDS and 82 (82%) agreed or strongly agreed to use it if given access. Perceptions about the accuracy of the AI-based CDS were largely positive, where 76 (76%) participants perceived their decisions on risk more accurate, while 6 (6%) somewhat disagreed with its accuracy.

For the automated CDS, 72 (70%) agreed or strongly agreed to using its recommendations, and 85 (83%) agreed or strongly agreed to using it if given access. Perceptions about the accuracy of the automated CDS were largely positive, where 75 (73%) participants perceived their decisions on risk were more accurate, while 3 (3%) somewhat disagreed with its accuracy ().

‎

Qualitative comments highlighted the overall usefulness of both CDS in providing comprehensive patient management (). Doctors believed that CDS had the potential to improve the efficiency of risk assessment and decision-making, particularly in primary care settings where patient volumes are high. While the AI-based CDS was perceived to be more accurate, the need to verify recommendations against clinical guidelines based on the patient’s actual condition was noted. The necessity for the AI-based CDS to offer explanations of its risk scores was also highlighted (, Doctor 014, Doctor 007). Automated input of patient data and the ability to work offline without an internet connection were identified as essential features for ease of use.

Discussion

Principal Results and Implications

We found that use of AI-based CDS significantly improved 10-year ASCVD risk assessment and decreased decision-making time, with mixed effects on patient management. Overall, participants’ perceptions about AI assistance were favorable. The observed improvement indicates potential for AI-based CDS to improve routine screening for CVD risk factors in resource-constrained settings. The number needed to treat for benefit associated with AI-based CDS was around 4, indicating that for every 4 patients where doctors use AI-based CDS as the “treatment,” there will be 1 additional correct risk classification as the “outcome” (number needed to treat=3.7, 95% CI 2.9-5.2). Assuming daily primary care visits of 30-100 patients per doctor (AP Susanto, PhD, et al; unpublished data, August 2024), with 12% requiring ASCVD risk assessment [], this amounts to an additional 1-3 accurate risk assessments per day, totaling 250-750 per year. Accordingly, these findings suggest the need for further research in a live clinical environment to determine whether improvements observed in the online study can be translated into real-world clinical settings.

For AI-based CDS to provide value, the decisions made by doctors must trigger some action or the CDS must provide actionable recommendations []. We measured patient management as an indicator of change in actionable care delivery, specifically the correct prescription of primary preventive medications (aspirin and statin), prescription of antihypertensive medications, and referral for advanced examinations. The findings indicated mixed effects on care delivery. Most importantly, AI-based CDS positively impacted the prescription of statins, showing that primary prevention of ASCVD could be improved by deploying AI tools to assist doctors. This is consistent with previous studies in Indonesia that identified gaps in awareness about the prescription of statins [], which, when prescribed at an appropriate intensity based on individual risk assessments, can reduce the 10-year incidence of ASCVD, achieve lipid target levels, and minimize medication side effects for long-term safety [].

Conversely, the effects of AI-based CDS in improving the prescription of aspirin and antihypertensive medications did not reach statistical significance. One possible explanation is that doctors may consider blood pressure independently of the ASCVD risk assessment and CDS recommendation when prescribing antihypertensives []. For aspirin, which had fewer cases of correct prescription, doctors appeared to ignore the CDS recommendations or did not see the guideline. Another possibility is a ceiling effect: the baseline correctness for these 3 measures in the control group (67.3% and 80.1%, respectively) was already substantially higher than for statins (35.3%), leaving less room for improvement. Consequently, doctors may have benefited less from the CDS when prescribing aspirin and antihypertensives, thereby limiting potential gains. Further research is needed to understand these effects.

We could not ascertain why AI-based CDS did not improve referral for advanced examinations. One possible explanation is that referrals were not included in the AI-based CDS recommendations, requiring doctors to infer the need for referral based on their knowledge of local access and clinical judgment. This is consistent with the ambiguity in the current national guidelines, which are silent about the referral of high-risk patients for advanced examination, possibly due to variations in local resources and policies. As such, this finding implies that the use of AI-based CDS, particularly for risk assessment, should be augmented with clear clinical consensus and local policy to enhance the actionability of recommendations to avoid mixed or negative effects on care delivery.

Delivering Fast and Acceptable AI-Based CDS

We found that use of AI-based CDS reduced decision-making time by 12% compared to the control group, even though doctors with AI assistance received additional information and confirmed the risk assessment provided by the system. While the reduction in decision-making time is statistically significant, the 10-second difference may not be clinically meaningful. Nevertheless, this finding implied that decision-making assisted by AI remained faster than heuristic risk assessment without CDS. This suggests that using AI-based CDS decreases decision time, despite requiring doctors to review and confirm the AI output.

The observed effects on decision-making time are consistent with previous studies of AI in primary care. For instance, a randomized controlled study of cataract screening in a remote area suggested faster diagnosis when supported by AI-based CDS compared to doctors alone (2.79 min vs 8.52 min, P<.001) []. In another study simulating AI-based CDS for the treatment of depression, 40% of doctors indicated that using CDS would save time []. Thus, AI-based CDS may be especially valuable in high-volume, resource-constrained settings where clinician time is scarce (AP Susanto, PhD, et al; unpublished data, August 2024).

Doctors generally viewed CDS for ASCVD risk assessment positively, but they raised concerns about its accuracy, guideline adherence, and capacity to address patients comprehensively. Similar concerns were observed with the use of AI-based risk assessment for delirium []. In our study, most doctors found AI-based CDS useful for decision-making, and their intention to use it was reinforced by perceptions of its high accuracy. Conversely, an AI system for sepsis prediction was poorly rated due to late, less useful alerts and a lack of transparency in its “black-box” results []. Doctors who perceive AI as beneficial are more likely to adopt it in clinical practice.

Safe Use of AI-Based CDS

Despite having access to highly accurate AI-based CDS, approximately 89 (29%) of patient cases were inaccurately assessed, as doctors failed to recognize accurate recommendations. This proportion appears to compare with 24 (24%) participants, who were neutral or somewhat disagreed with the accuracy of the AI in the TAM survey. Participant comments suggested that doctors sought explanations when their risk estimates diverged from CDS recommendations (). In such cases, doctors might have relied on their own risk estimates due to the AI’s lack of explanation.

For future development and validation, one way to increase doctors’ confidence and performance is for AI-based CDS to provide an explanation as demonstrated by the mock-up in . Explanations could be model-agnostic and local, displaying risk factors, input variables, their marginal contributions, and risk scores, rather than only providing on risk level recommendations. Previous studies demonstrated that model-agnostic local explanations, which include individual risk factors, can help doctors comprehend, trust, and explain outputs to patients []. As doctors prefer simple visualizations and relevant clinical reasoning, AI should empower them to offer meaningful explanations to patients, facilitating shared decision-making.

‎

Our results suggest that doctors need support through relevant clinical guidelines and training to make informed decisions about using and evaluating AI recommendations. Although participants had access to clinical guidelines, and comments emphasized their importance (), our findings are consistent with a previous study of prescribing decision support with 102 participants, where less accurate recommendations were accepted despite the availability of guidelines for verification []. Differently from high-resource settings, where verification can involve advanced examinations or medical specialists, in resource-constrained settings, verification mechanisms may be limited to simple tools such as guidelines. Here, doctors must proceed with the final decision and treatment assisted by AI-based risk assessment alone. Preemptive training is crucial for safe CDS use, especially with AI, as each tool will operate differently. Doctors need to understand what inputs the AI considers or excludes and be aware of how incorrect use could harm patients []. This study therefore highlights the need for AI developers and users to address safety considerations in AI-based CDS for resource-limited settings, to enhance global health equity rather than introduce new risks or burdens.

Strengths and Limitations

This is the first randomized controlled study to evaluate the effects of AI on ASCVD risk assessment and patient management for primary prevention in a resource-constrained setting. Participants were instructed to perform the risk assessment as they would in a real outpatient setting, exercising their best clinical judgment. The awareness of completing clinical vignettes might have influenced doctors to perform at an enhanced level compared to real-world practice and overestimate improvement []. Conversely, doctors could assume that any inaccurate decision carried no accountable consequence.

We specifically designed the cases for resource-constrained clinical settings where there is access to information on traditional risk factors but limited access to advanced examinations such as computed tomography calcium scores, lipoprotein (a), and medical specialists. Our findings can therefore be generalized to similar settings. However, the spectrum within resource-constrained settings varies both horizontally across different dimensions and vertically within a single dimension []. For example, within the dimension of limited access to advanced examinations, even traditional risk factors like high-density lipoprotein cholesterol tests may be unavailable in some settings. Efforts were made to ensure that the cases were representative of the clinical context; however, this may not reflect the true rate of CVD risk in the population. Although the minimum sample size was determined a priori and was adequate, recruitment through the Indonesian Heart Association at 5 training centers may have led to a participant group more predisposed to research involvement than the broader medical community. Therefore, the findings may not be generalizable to the broader population of Indonesian GPs, who may be older, have different training backgrounds, and practice in more isolated community health centers ().

Additionally, we used a conceptual prototype of an AI-based CDS that accounted for REFs based on expert consensus, which may not reflect the true performance difference between an actual AI-based tool and the existing risk calculator. Actual AI-based tools have dynamic, rather than fixed, performance as they are continually trained on new data. In addition, the conceptual prototype did not account for practical challenges such as input data quality and model transportability across settings, which may limit the generalizability of the results to real-world settings. Nevertheless, this study has demonstrated the potential of AI-based CDS to improve risk assessment by accounting for patient-level precision beyond traditional risk factors and sets the foundation for further studies to test AI-based tools ahead of clinical implementation.

Conclusions

Improvements in ASCVD risk assessment with a conceptual prototype demonstrate that use of AI can support better and faster decision-making in resource-constrained settings and are moderated by doctors’ positive perceptions.

SANTA CLARA, Calif., Nov. 25, 2025 – Intel Corporation today announced that John Pitzer, corporate vice president, Global Treasury and Investor Relations, will participate in fireside chats to discuss Intel’s business and strategy at the following investor events:

• On Dec. 3 at 1:15 p.m. PT: UBS Global Technology and AI Conference.
• On Dec. 10 at 11:35 a.m. PT: Barclays Global Technology Conference. 

Live webcasts and replays can be accessed publicly on Intel’s Investor Relations website at intc.com.

Intel’s participation, speakers and schedule are subject to change.

Note: This story has been updated to reflect a revised date and time for the upcoming UBS Global Technology and AI Conference fireside chat.

About Intel

Intel (Nasdaq: INTC) designs and manufactures advanced semiconductors that connect and power the modern world. Every day, our engineers create new technologies that enhance and shape the future of computing to enable new possibilities for every customer we serve. Learn more at intel.com. 

© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries. Other names and brands may be claimed as the property of others.

Contacts:

Investor Relations
investor.relations@intel.com

Sophie Metzger
Media Relations
sophie.metzger@intel.com

Released Nov 25, 2025 • 4:25 PM EST

Good morning and welcome back to FirstFT Asia. In today’s newsletter:

---

Volkswagen has said it can produce an electric vehicle entirely made in China for half the cost of doing so elsewhere, as the German automaker fights to reclaim its share in the world’s biggest market.

What to know: Europe’s largest carmaker said yesterday that, following a series of investments in China, it can for the first time develop new cars outside Germany. VW is preparing to release about 30 EV models in China over the next five years, in a bet on localised research and development. The carmaker said that, compared with the 2023 production costs for EVs in Germany, the cost for some models in China had been reduced by as much as 50 per cent because of supply chain efficiencies.

VW was for many years the biggest carmaker in China but it has lost market share as a wave of innovative Chinese EV rivals won over consumers.

VW’s China plans: The group is in discussions about increasing exports of Chinese-made cars as well as applying China breakthroughs throughout its global operations, said executives. VW is among a clutch of European carmakers attempting to replicate the speed of vehicle development in China by hiring local engineers, collaborating with partners there and by tearing down vehicles made by BYD and other newer rivals. Read more about VW’s efforts to survive in China.

Here’s what else we’re keeping tabs on today:

• Economic data: Australia publishes October inflation data.

• UK economy: Chancellor Rachel Reeves unveils a massive tax-raising Budget in what is perhaps the defining event of this parliament.

Join senior Financial Times editors on December 3 for a discussion on what will shape the year ahead. Register for free today.

Five more top stories

1. Donald Trump said he would dispatch special envoy Steve Witkoff to meet Russian leader Vladimir Putin in Moscow in another push to secure a peace plan to end the Kremlin’s war in Ukraine. The US president said his negotiators had made “tremendous progress” in ending the war — but he would wait until a deal was within reach before meeting Ukraine’s leader Volodymyr Zelenskyy and Putin. Here’s the latest on the peace talks.

2. Samsung chair Lee Jae-yong hosted Mukesh Ambani, Asia’s richest man and chair of India’s Reliance Industries, for talks over deepening the relationship between two of the region’s most powerful conglomerates. Lee has been holding a series of high-profile meetings as he seeks to keep his company in a strong position for the global AI build-out.

3. Nvidia shares fell sharply yesterday on fears that Google is gaining ground in AI, erasing $150bn in market value from the chipmaker. Google last week released Gemini 3, its latest large language model, which is considered to have leapfrogged OpenAI’s ChatGPT. Analysts likened Gemini 3’s impact to the disruption triggered by Chinese start-up DeepSeek.

4. Indonesia is investigating radioactive contamination at an industrial zone after US and Dutch authorities found unusual levels of radiation in some of the south-east Asian country’s biggest exports. A total of 22 factories were affected, including facilities that process shrimp and make footwear for Nike and Adidas. Here are more details.

5. A series of grim official data released in the US has deepened concerns about the health of the world’s largest economy. Signs of weakness in retail sales and consumer confidence released yesterday suggest Americans are pulling back on spending amid an affordability crisis that has raised pressure on Trump.

• ‘Scraping for crumbs’: High prices of food, rents and healthcare are forcing lower-income Americans to cut back on necessities just as the Trump administration curbs government supports.

The Big Read

Ireland’s location on Europe’s western fringe has made it pivotal to global communications. But with a small navy and limited intelligence sharing, some say the military-neutral EU member is incapable of protecting itself and the transatlantic data cables running through its territory. Read more on what makes the island nation a weak link in EU defence.

We’re also reading . . . 

• Rare earths: A small Australian town best known for its giant Elvis Presley festival is gaining global significance in the race to break China’s control of rare earths.

• Jair Bolsonaro: A Brazilian supreme court judge has ordered the ex-president to begin a 27-year jail sentence following his conviction for coup plotting.

• Global depopulation: What would it feel like to live in a shrinking world? One view is that it would be bleak. Sarah O’Connor is more optimistic.

Chart of the day

The Fractured Age, a new book by Neil Shearing of Capital Economics, argues that the world economy will divide between a US-centred bloc and a China-centred one. But which country will come out ahead? Martin Wolf examines this question in his latest column.

Take a break from the news . . .

HTSI horology expert Nick Foulkes picks his top timepieces of 2025.

Net revenue and EPS results
HP Inc. and its subsidiaries (“HP”) announced fiscal 2025 net revenue of $55.3 billion, up 3.2% (up 3.7% in constant currency) from the prior-year period.
 
Fiscal 2025 GAAP diluted net EPS was $2.65, down from $2.81 in the prior-year. Fiscal 2025 non-GAAP diluted net EPS was $3.12, down from $3.43 in the prior-year period. Fiscal 2025 non-GAAP net earnings and non-GAAP diluted net EPS exclude after-tax adjustments of $443 million, or $0.47 per diluted share, related to restructuring and other charges, acquisition and divestiture charges, amortization of intangible assets, certain litigation charges (benefits), net, non-operating retirement-related credits, tax adjustments, and the related tax impact on these items.
 
Fourth quarter net revenue was $14.6 billion, up 4.2% (up 3.8% in constant currency) from the prior-year period.
 
Fourth quarter GAAP diluted net EPS was $0.84, down from $0.93 in the prior-year period and within the previously provided outlook of $0.75 to $0.85. Fourth quarter non-GAAP diluted net EPS was $0.93, down from $0.96 in the prior-year period and within the previously provided outlook of $0.87 to $0.97. Fourth quarter non-GAAP net earnings and non-GAAP diluted net EPS excludes after-tax adjustments of $82 million, or $0.09 per diluted share, related to restructuring and other charges, acquisition and divestiture charges, amortization of intangible assets, certain litigation charges (benefits), net, tax adjustments, and the related tax impact on these items.
 
“HP’s strategy to lead the Future of Work continues to deliver strong performance, marked by our sixth consecutive quarter of revenue growth,” said Enrique Lores, President and CEO of HP Inc. “Our FY25 results reinforce the power of our portfolio and the strength of our team in a dynamic environment. As we accelerate innovation across AI-powered devices to drive productivity, security, and flexibility for our customers, our focus for FY26 is on disciplined execution. We are committed to driving measurable results – ensuring that our plans translate into long-term value for our shareholders.”
 
“Our FY25 results reflect solid execution in an evolving environment, where we drove strong profit improvement in the back half of the year and returned $1.9 billion dollars to shareholders,” said Karen Parkhill, CFO, HP Inc. “Looking forward, we are taking decisive actions to mitigate recent cost headwinds and are investing in AI-enabled initiatives to accelerate product innovation, improve customer satisfaction, and boost productivity. We are confident these actions will strengthen our foundation and position us for long-term growth.”
 
Asset management
HP generated $3.7 billion in net cash provided by operating activities and $2.9 billion of free cash flow in fiscal 2025. Free cash flow includes net cash provided by operating activities of $3.7 billion adjusted for net investments in leases from integrated financing of $131 million and net investments in property, plant, equipment and purchased intangibles of $897 million. HP utilized $850 million of cash during fiscal 2025 to repurchase approximately 29.5 million shares of common stock in the open market. When combined with the $1.1 billion of cash used to pay dividends, HP returned 66% of its free cash flow to shareholders in fiscal 2025.
 
HP’s net cash provided by operating activities in the fourth quarter of fiscal 2025 was $1.6 billion. Accounts receivable ended the quarter at $5.7 billion, up 2 days quarter over quarter to 35 days. Inventory ended the quarter at $8.5 billion, down 2 days quarter over quarter to 66 days. Accounts payable ended the quarter at $18.1 billion, up 1 day quarter over quarter to 139 days.
 
HP generated $1.5 billion of free cash flow in the fourth quarter. Free cash flow includes net cash provided by operating activities of $1.6 billion adjusted for net investments in leases from integrated financing of $60 million and net investments in property, plant and equipment of $197 million.
 
HP’s dividend payment of $0.2894 per share in the fourth quarter resulted in cash usage of $270 million. HP also utilized $500 million of cash during the quarter to repurchase approximately 18.3 million shares of common stock in the open market. HP exited the quarter with $3.7 billion in gross cash, which includes cash and cash equivalents of $3.7 billion, restricted cash of $15 million, cash held for sale of $8 million, and short-term investments of $3 million included in other current assets. Restricted cash is related to amounts collected and held on behalf of a third party for trade receivables previously sold.
 
The HP board of directors has declared a quarterly cash dividend of $0.30 per share on the company’s common stock, payable on January 2, 2026 to stockholders of record as of the close of business on December 11, 2025. This is the first dividend of HP’s 2026 fiscal year.
 
Fiscal 2025 fourth quarter segment results

• Personal Systems net revenue was $10.4 billion, up 8% year over year (up 7% in constant currency) with a 5.8% operating margin. Consumer PS net revenue was up 10% and Commercial PS net revenue was up 7%. Total units were up 7% with Consumer PS units up 8% and Commercial PS units up 7%.

• Printing net revenue was $4.3 billion, down 4% year over year (down 4% in constant currency) with an 18.9% operating margin. Consumer Printing net revenue was down 9% and Commercial Printing net revenue was down 4%. Supplies net revenue was down 4% (down 3% in constant currency). Total hardware units were down 12%, with both Consumer and Commercial Printing units reflecting similar declines.

 
Outlook
For the fiscal 2026 first quarter, HP estimates GAAP diluted net EPS to be in the range of $0.58 to $0.66 and non-GAAP diluted net EPS to be in the range of $0.73 to $0.81. Fiscal 2026 first quarter non-GAAP diluted net EPS estimates exclude $0.15 per diluted share, primarily related to restructuring and other charges, acquisition and divestiture charges, amortization of intangible assets, non-operating retirement-related credits, tax adjustments, and the related tax impact on these items.
 
For fiscal 2026, HP estimates GAAP diluted net EPS to be in the range of $2.47 to $2.77 and non-GAAP diluted net EPS to be in the range of $2.90 to $3.20. Fiscal 2026 non-GAAP diluted net EPS estimates exclude $0.43 per diluted share, primarily related to restructuring and other charges, acquisition and divestiture charges, amortization of intangible assets, non-operating retirement-related credits, tax adjustments, and the related tax impact on these items. For fiscal 2026, HP anticipates generating free cash flow in the range of $2.8 to $3.0 billion.  HP’s outlook reflects the added cost driven by the current U.S. trade-related regulations in place, and associated mitigations. 
 
More information on HP’s earnings, including additional financial analysis and an earnings overview presentation, is available on HP’s Investor Relations website at investor.hp.com.
 
HP’s FY25 Q4 earnings conference call is accessible via audio webcast at www.hp.com/investor/2025Q4Webcast.
 
Company-wide initiative announced in November 2025
Today, HP Inc. announced a company-wide initiative (“fiscal 2026 plan”) to drive customer satisfaction, product innovation, and productivity through artificial intelligence adoption and enablement. The company estimates that these actions will result in gross run rate savings of approximately $1 billion by the end of fiscal 2028. The company estimates that it will incur approximately $650 million in labor and non-labor costs related to restructuring and other charges, with approximately $250 million in fiscal 2026. The company expects to reduce gross global headcount by approximately 4,000-6,000 employees. These actions are expected to be completed by the end of fiscal 2028.
 
About HP Inc.
HP Inc. (NYSE:HPQ) is a global technology leader redefining the Future of Work. Operating in more than 180 countries, HP delivers innovative and AI-powered devices, software, services and subscriptions that drive business growth and professional fulfillment. For more information, please visit: HP.com.
 
Use of non-GAAP financial information
To supplement HP’s consolidated condensed financial statements presented on a generally accepted accounting principles (“GAAP”) basis, HP provides net revenue on a constant currency basis, non-GAAP total operating expense, non-GAAP operating profit, non-GAAP operating margin, non-GAAP other income and expenses, non-GAAP tax rate, non-GAAP net earnings, non-GAAP diluted net EPS, free cash flow, gross cash and net cash (debt) financial measures. HP also provides forecasts of non-GAAP diluted net EPS and free cash flow. Reconciliations of these non-GAAP financial measures to the most directly comparable GAAP financial measures are included in the tables below or elsewhere in the materials accompanying this news release. In addition, an explanation of the ways in which HP’s management uses these non-GAAP measures to evaluate its business, the substance behind HP’s decision to use these non-GAAP measures, the material limitations associated with the use of these non-GAAP measures, the manner in which HP’s management compensates for those limitations, and the substantive reasons why HP’s management believes that these non-GAAP measures provide useful information to investors is included under “Use of non-GAAP financial measures” after the tables below. This additional non-GAAP financial information is not meant to be considered in isolation or as a substitute for net revenue, operating expense, operating profit, operating margin, other income and expenses, tax rate, net earnings, diluted net EPS, cash provided by operating activities or cash, cash equivalents, and restricted cash prepared in accordance with GAAP.
 
Forward-looking statements
This document contains forward-looking statements based on current expectations and assumptions that involve risks and uncertainties. If the risks or uncertainties ever materialize or the assumptions prove incorrect, they could affect the business and results of operations of HP Inc. and its consolidated subsidiaries which may differ materially from those expressed or implied by such forward-looking statements and assumptions.
 
All statements other than statements of historical fact are statements that could be deemed forward-looking statements, including, but not limited to, projections of net revenue, margins, expenses, effective tax rates, net earnings, net earnings per share, cash flows, benefit plan funding, deferred taxes, share repurchases, foreign currency exchange rates or other financial items; any projections of the amount, timing or impact of cost savings or restructuring and other charges, planned structural cost reductions and productivity initiatives; any statements of the plans, strategies and objectives of management for future operations, including, but not limited to, our business model and transformation, our sustainability goals, our go-to-market strategy, the execution of restructuring plans and any resulting cost savings (including the fiscal 2023 plan and the fiscal 2026 plan), net revenue or profitability improvements or other financial impacts; any statements concerning the expected development, demand, performance, market share or competitive performance relating to products or services; any statements concerning potential supply constraints, component shortages, manufacturing disruptions or logistics challenges; any statements regarding current or future macroeconomic trends or events, including global trade policies, and the impact of those trends and events on HP and its financial performance; any statements regarding pending investigations, claims, disputes or other litigation matters; any statements of expectation or belief as to the timing and expected benefits of acquisitions and other business combination and investment transactions; and any statements of assumptions underlying any of the foregoing.  Forward-looking statements can also generally be identified by words such as “future,” “anticipates,” “believes,” “estimates,” “expects,” “intends,” “plans,” “predicts,” “projects,” “will,” “would,” “could,” “can,” “may,” and similar terms.
 
Risks, uncertainties and assumptions that could affect our business and results of operations include factors relating to HP’s ability to execute on its strategic plans, including the previously announced initiatives, business model changes and transformation; the development and transition of new products and services and the enhancement of existing products and services to meet evolving customer needs and respond to emerging technological trends, including artificial intelligence; the use of artificial intelligence; the impact of macroeconomic and geopolitical trends, changes and events, including global trade policies, the ongoing military conflict in Ukraine, continued instability in the Middle East or tensions in the Taiwan Strait and South China Sea and the regional and global ramifications of these events; volatility in global capital markets and foreign currency, changes in benchmark interest rates, the effects of inflation and instability of financial institutions; risks associated with HP’s international operations and the effects of business disruption events, including those resulting from climate change; the need to manage (and reliance on) third-party suppliers, including with respect to supply constraints and component shortages, and the need to manage HP’s global, multi-tier distribution network and potential misuse of pricing programs by HP’s channel partners, adapt to new or changing marketplaces and effectively deliver HP’s services; the execution and performance of contracts by HP and its suppliers, customers, clients and partners, including logistical challenges with respect to such execution and performance; the competitive pressures faced by HP’s businesses; the impact of third-party claims of IP infringement; successfully innovating, developing and executing HP’s go-to-market strategy, including online, omnichannel and contractual sales, in an evolving distribution, reseller and customer landscape; successfully competing and maintaining the value proposition of HP’s products, including supplies and services; challenges to HP’s ability to accurately forecast inventories, demand and pricing, which may be due to HP’s multi-tiered channel, sales of HP’s products to unauthorized resellers or unauthorized resale of HP’s products or our uneven sales cycle; the hiring and retention of key employees; the results of our restructuring plans (including the fiscal 2023 plan and the fiscal 2026 plan), including estimates and assumptions related to the cost (including any possible disruption of HP’s business) and the anticipated benefits of our restructuring plans; the protection of HP’s intellectual property assets, including intellectual property licensed from third parties; disruptions in operations from system security risks, data protection breaches, or cyberattacks; HP’s ability to maintain its credit rating, satisfy its debt obligations and complete any contemplated share repurchases, other capital return programs or other strategic transactions; changes in estimates and assumptions HP makes in connection with the preparation of its financial statements; the impact of changes to federal, state, local and foreign laws and regulations, including environmental regulations and tax laws; integration and other risks associated with business combination and investment transactions; our aspirations related to environmental and societal matters; potential impacts, liabilities and costs from pending or potential investigations, claims and disputes; the effectiveness of our internal control over financial reporting; and other risks that are described in HP’s Annual Report on Form 10-K for the fiscal year ended October 31, 2024 and HP’s other filings with the Securities and Exchange Commission (“SEC”). HP’s fiscal 2023 plan included HP’s efforts to take advantage of future growth opportunities, including but not limited to, investments to drive growth, investments in our people, improving product mix, driving structural cost savings and other productivity measures. HP’s fiscal 2026 plan includes HP’s efforts to drive customer satisfaction, product innovation, and productivity through artificial intelligence adoption and enablement, and cost savings associated with the fiscal 2026 plan represent gross reductions in costs from these restructuring plans. Structural cost savings represent gross reductions in costs driven by operational efficiency, digital transformation, and portfolio optimization. These initiatives include but are not limited to workforce reductions, platform simplification, programs consolidation and productivity measures undertaken by HP, which HP expects to be sustainable in the longer-term. These structural cost savings are net of any new recurring costs resulting from these initiatives and exclude one-time investments to generate such savings. HP’s expectations on the longer-term sustainability of such structural cost savings are based on its current business operations and market dynamics and could be significantly impacted by various factors, including but not limited to HP’s evolving business models, future investment decisions, market environment and technology landscape.
 
As in prior periods, the financial information set forth in this document, including any tax-related items, reflects estimates based on information available at this time. While HP believes these estimates to be reasonable, these amounts could differ materially from reported amounts in HP’s Annual Report on Form 10-K for the fiscal year ending October 31, 2025 and October 31, 2026, Quarterly Report on Form 10-Q for the fiscal quarter ending January 31, 2026 and HP’s other filings with the SEC. The forward-looking statements in this document are made as of the date of this document and HP assumes no obligation and does not intend to update these forward-looking statements. 
 
HP’s Investor Relations website at investor.hp.com contains a significant amount of information about HP, including financial and other information for investors. HP encourages investors to visit its website from time to time, as information is updated, and new information is posted. The content of HP’s website is not incorporated by reference into this document or in any other report or document HP files with the SEC, and any references to HP’s website are intended to be inactive textual references only. 
 
 
Available in PDF format — including all tables

As AI tools advance, multinational employers are grappling with a growing patchwork of governance frameworks. Notably in the Asia-Pacific (APAC) region, Japan, Singapore, India, and Australia are making moves that encourage AI use while seeking to curb the risks. Businesses operating in these locations should stay on top of recent developments and create an action plan to address the nuances in each applicable location. Here’s a breakdown of new developments in this region and what you can do to prepare as AI regulatory frameworks continue to take shape.

Japan Focuses on Innovation

Earlier this year, the legislature in Japan passed a bill that encourages research and development – as well as advancements – in AI technology and commits to international leadership and cooperation. The new law sets expectations for businesses to cooperate with key principles for AI governance, including transparency and risk mitigation. Research institutions are expected to work on AI development and train skilled workers, and citizens are encouraged to learn more about AI. There are no sanctions for noncompliance. Rather, the focus is on voluntary collaboration and making strides in this realm.

Moreover, the law charges Japan’s national government with creating AI policies. Specifically, it establishes an Artificial Intelligence Strategy Headquarters within the Cabinet, which will oversee the new Artificial Intelligence Basic Plan and coordinate all AI policy. Local governments are responsible for implementing a regional plan.

Action Items: Businesses operating in Japan should be encouraged by the innovation-focused approach. Still, you should review (or develop) your AI policies for transparency and alignment with the new guidance. You should also consider actively participating in voluntary initiatives that support responsible AI development and use.

Singapore Highlights Testing and Sector-Specific Compliance

Balancing innovation with effective risk management is also a priority in Singapore. The government published its first Model AI Governance Framework in 2019 – which established principles for ethical AI use – and has invested millions in AI efforts. This year, through AI Verify Foundation, Singapore launched the Global AI Assurance Pilot, which served as a sandbox to test Generative AI applications and establish international standards. The purpose was “to help codify emerging norms and best practices around technical testing of GenAI applications.”

In terms of regulations, Singapore has chosen not to pursue a comprehensive AI statute. Instead, it continues to follow a sector-specific regulatory model that addresses risks through existing frameworks for finance, healthcare, employment, and other compliance areas.

Action Items: Explore AI Verify Foundation initiatives and review the pilot report. For ongoing compliance, continue to ensure AI practices are aligned with sector-specific regulations.

India Issues Robust Guidelines

India released its AI Governance Guidelines in 2025, emphasizing safety, trust, and flexibility. Specifically, the seven core principles are:

1. Trust is the Foundation
2. People First
3. Innovation over Restraint
4. Fairness and Equity
5. Accountability
6. Understandable by Design
7. Safety, Resilience, and Sustainability

The guidelines, which you can learn about in more detail here, make six key recommendations, including education and skills training, as well as policies that support innovation, mitigate risks, and encourage compliance through transparency and voluntary measures.

The guidelines also recommend setting up an AI Governance Group, which would be supported by a Technology and Policy Expert Committee – and enabling the AI Safety Institute to provide technical expertise on trust and safety issues.

As in most other countries, the recommendations also call for sector-specific regulations to continue. For example, the Digital Personal Data Protection Act (DPDP Act), India’s first comprehensive data privacy legislation regulating digital personal data processing, governs the use of personal data to train AI models.

Action Items: Review AI practices and assess how they measure up to India’s core principles, particularly on fairness, accountability, and transparency. Audit processes to ensure compliance with existing laws like the DPDP Act and other rules that impact your business.

Australia Shifts Gears

In 2024, Australia initially proposed mandatory guardrails for AI use in high-risk settings and released its Voluntary AI Safety Standard to serve as guidance while the requirements were pending.

The country has since changed its direction from aiming to roll out a stricter EU-type AI framework to one that looks more like other APAC nations, focused on safe and responsible AI use, as well as innovation and compliance with existing laws and regulations.

Last month, Australia’s National AI Centre issued new Guidance for AI Adoption, which replaces the prior voluntary standard and sets out six practices for responsible AI governance:

1. Decide who is accountable 
2. Understand impacts and plan accordingly 
3. Measure and manage risks
4. Share essential information
5. Test and monitor
6. Maintain human control

To learn more, you can visit the government’s resource page.

Action Items: Get familiar with the new Guidance for AI Adoption and the accompanying resources. Consider aligning internal AI practices with the six recommended governance steps – and continue to monitor developments as Australia shapes its approach to AI governance.

Key Takeaways for Multinational Businesses

• Global AI Regulations Are Evolving Quickly: While Japan, Singapore, India, and Australia have all introduced frameworks that balance innovation with risk mitigation, the European Union has taken a stricter approach and leading European companies are pushing back. You’ll want to track developments closely.
• No One-Size-Fits-All Plan: Carefully review the applicable countries’ approach to voluntary vs. mandatory guidance and sector-specific rules vs. comprehensive AI statutes.
• Themes Are Emerging: Although each approach is different, goals like safety, transparency, accountability, and human oversight are common.
• Existing Laws Still Apply: AI regulatory frameworks generally do not replace compliance obligations under data privacy, finance, healthcare, employment, and other laws that may apply when AI tools are used. Be sure that your AI, Tech, Legal, HR, and other teams coordinate compliance efforts.
• Privacy Rules Still Shape AI Use: Because AI is being handled through existing privacy frameworks, companies should watch how regulators interpret transparency and data use in AI contexts. These factors directly affect how AI models are trained and how they generate outputs.
• Sector-Specific Rules May Affect Certain AI Use Cases More Than Others: Given the region’s reliance on sector-based oversight, areas like recruitment, financial services, and healthcare may face closer scrutiny when AI tools are used in those contexts.
• Navigating AI Governance Can Be Challenging: Having a trusted legal or consulting partner can provide valuable guidance, ensuring compliance and aligning AI practices with business goals. From setting up policies to managing compliance, a partner experienced in AI governance can help prevent costly missteps.

Conclusion

We will continue to monitor legal changes affecting multinational companies, so make sure you are subscribed to Fisher Phillips’ Insight System to receive the latest updates directly to your inbox. If you have questions, contact your Fisher Phillips attorney, the authors of this Insight, or any attorney in our International Practice Group or our AI, Data, and Analytics Practice Group.

NEW YORK–(BUSINESS WIRE)–
UiPath, Inc. (NYSE: PATH), a global leader in agentic automation, today announced that Ashim Gupta, Chief Operating Officer and Chief Financial Officer, will participate in a fireside chat at the Barclays 23rd Annual Global Technology Conference, to be held at the Palace Hotel in San Francisco, CA on Wednesday, December 10th at 11:00 am PT (2:00 pm ET).

The presentation will be available via live audio webcast and archived replay on the Investor Relations section of the Company’s website (https://ir.uipath.com).

About UiPath

UiPath (NYSE: PATH) is a global leader in agentic automation, empowering enterprises to harness the full potential of AI agents to autonomously execute and optimize complex business processes. The UiPath Platform™ uniquely combines controlled agency, developer flexibility, and seamless integration to help organizations scale agentic automation safely and confidently. Committed to security, governance, and interoperability, UiPath supports enterprises as they transition into a future where automation delivers on the full potential of AI to transform industries. For more information, visit www.uipath.com.

View source version on businesswire.com: https://www.businesswire.com/news/home/20251125382239/en/

Investor Relations Contact

Allise Furlani

investor.relations@uipath.com

UiPath

Media Contact

pr@uipath.com

UiPath

Source: UiPath, Inc.

Released November 25, 2025