Category: 3. Business

Introduction

Gait monitoring via digital health technologies, such as wearable inertial sensors, continues to grow in prevalence and acceptance. Especially given the ability of these digital health technologies to enable remote and continuous monitoring, these quantitative data have provided key insights for clinicians, including the ability to monitor or diagnose a variety of diseases and populations []. This increased acceptance is highlighted by the recent approval of the 95th percentile of gait speed as a primary end point in Duchenne Muscular Dystrophy trials []. Gait has also been referred to as a “functional sixth vital sign” [,], and, along with previous research showing relationships with all-cause mortality, disability status, age, and more [-], highlights the importance of continuous gait monitoring. With patient centricity as the main objective, remote and continuous gait monitoring provides opportunities for trial sponsors to consider decentralized clinical trials, which can reduce patient burden and increase patient diversity [].

Researchers have developed algorithms to extract spatial and temporal gait metrics using data collected by sensors at various body locations, such as the wrist, hip, and lower trunk, which has enhanced the understanding of human locomotion dynamics [-].

Beyond the lumbar and wrist, numerous other sensor placements have been investigated. Foot- and shank-mounted inertial sensors have long been regarded as reference-standard placements, as they directly capture ground contact events and enable accurate stride length and speed estimation [,]. However, their use in large-scale or decentralized studies is limited by setup burden, footwear dependence, and reduced patient compliance compared with lumbar or wrist-worn devices. More recently, innovative placements such as head- and ear-worn sensors have been explored, leveraging hearing aids, earbuds, or headbands to unobtrusively capture gait dynamics [-]. These approaches demonstrate the diversity of wearable configurations under study, though many face challenges in patient compliance, real-world wearability, or algorithm generalizability.

To increase overall compliance and reduce participant burden, it is of great interest to use a single sensor to monitor bilateral gait. The lumbar location is the most popular choice, though it has its drawbacks. Namely, accuracy can sometimes suffer due to its distal location relative to the feet and ground contact points. This is especially true for spatial metrics, which rely on modeling steps as pendulum swings [,], as opposed to direct integration (which has its own numerical problems) or the location and trajectory estimation methods typically used for sensors placed on the feet or legs. This lower accuracy is generally the result of a relatively constant bias in spatial metric estimation, such as stride length and gait speed [,,], even when algorithm precision remains high. Some previous works have addressed this drawback with personalized constants requiring calibration []; however, this potentially adds significant participant burden and limits the ability to conduct fully remote trials. Other remedies include using a relation between step frequency and step length [], machine/deep learning approaches, or employing different sensor locations []. However, these may suffer from a lack of generalizability, as nonphysical models or relationships were employed in the estimation procedure. Therefore, it is of interest to address these drawbacks within the current physical inverted pendulum model and improve accuracy.

While the lumbar site has technical limitations due to reliance on pendulum modeling for spatial metrics, it also offers practical advantages. Compared with placements such as the feet, chest, or head, the lumbar sensor is generally easier for participants to wear consistently, less intrusive in daily activities, and less affected by soft tissue or motion artifacts. Previous work has highlighted the lumbar site as a practical compromise between accuracy and wearability, supporting its use in both older adults and patient populations where long-term compliance is critical [,]. These attributes, coupled with its ability to capture bilateral gait from a single device, make the lumbar location particularly appealing for large-scale monitoring and decentralized trial settings, thereby reinforcing the rationale for optimizing algorithms at this site.

While several algorithms have been developed to estimate gait parameters from sensors placed at various body locations, the options for a comprehensive solution that can be implemented in clinical trials remain limited. One such option is our open-source SciKit Digital Health (SKDH) [], a Python package that includes various digital health algorithms (eg, sleep, physical activity, gait), data ingestion tools, and preprocessing methods. The gait module in SKDH, that is, SKDH-gait, evolved from GaitPy, a stand-alone open-source package [] that provided an interface for estimating gait metrics within a fixed framework. While the majority of the gait implementation remained the same, additional options were added, such as a dynamic relation between step frequency and wavelet scale [], which has already improved overall accuracy. Validation results for both the original GaitPy [] and the SKDH-gait implementation [,] have been previously published. However, a direct comparison between the 2 versions has never been officially reported. Another issue with these implementations is that a population-specific parameter specification is required, depending on factors such as age, which can create an additional operational burden. Other gait algorithms include the recently published Mobilise-D gait algorithm [,] and the gaitmap project []. While the gaitmap project focuses primarily on foot-placed devices—requiring 2 sensors for bilateral assessment in addition to the drawbacks noted above—the Mobilise-D gait algorithm utilizes a single lumbar device and has been released as an open-source package on GitHub, MobGap []. The Mobilise-D gait algorithm, therefore, occupies a similar space to SKDH’s gait algorithm as an alternative processing package.

As a matter of fact, at-home gait monitoring has already been implemented in clinical trials, for example, in cancer cachexia phase 1 and 2 studies, as end points designed to support clinical interpretation and decision-making in decentralized trials [,]. However, improving accuracy has concrete benefits for clinical translation. In patient populations, small but clinically meaningful changes in gait—such as reductions in stride length in multiple sclerosis [] or subtle slowing that predicts conversion to Parkinson disease []—can be obscured by algorithmic bias or variability, as even small gait deviations can be clinically meaningful. By reducing measurement error, more accurate algorithms increase sensitivity to detect disease progression and treatment response. In clinical trials, this enhanced sensitivity translates into stronger statistical power and potentially smaller required sample sizes. Finally, in the context of real-world monitoring and regulatory acceptance, reliable gait metrics are essential to ensure that longitudinal changes reflect true functional decline or improvement rather than artifacts of the analytic method. This aligns with Food and Drug Administration (FDA) guidance emphasizing the need for trustworthy digital end points in patient-focused drug development [].

To address the technical challenges mentioned above, reduce bias in the estimation of spatial metrics, and improve overall accuracy, the SKDH-gait algorithm has been overhauled and updated. These updates are significant enough to require new validation beyond what has been previously presented. This work aims to provide a step-by-step overview of the improvements and their validation using data collected from healthy adults and pediatric populations. We first consider each of the constituent blocks of the gait algorithm and examine areas requiring improvement, and then assess these proposed changes. Next, we combine all individual improvements in the full algorithm implementation and validate the updated approach and its effect on the final gait metrics compared with a reference system. Finally, we assess the test-retest reliability of the gait metrics. This paper serves as the first part of the validation work, beginning with healthy participants before transitioning to patient populations in future studies.

Methods

Participants and Procedures

Data used in this work for validating the gait algorithm were drawn from the In-Clinic and Natural Gait Observations (ICANGO) [] and the Monitoring Activity and Gait in Children (MAGIC) [] studies completed at the Pfizer Innovation Research Lab in Cambridge, Massachusetts.

In MAGIC, a total of 40 healthy ambulatory participants aged 3-17 years were enrolled. Participants were recruited equally from 3 independent age groups as follows: 3-5 years, 6-11 years, and 12-17 years. Informed consent was obtained from the parents or legally acceptable representatives of all participants. Assent was obtained in an age-appropriate manner (verbal assent for children 3-5 years; written assent for children 6-17 years). The study was conducted at the Pfizer Innovation Research Lab in Cambridge, Massachusetts, from 2021 to 2022.

In ICANGO, 2 cohorts of 20 healthy individuals each (40 participants total) over 18 years old (range 25-82 years) were enrolled. Participants did not have any significant health problems, as determined by a medically qualified study doctor during initial intake.

Details of the 2 studies have been previously published [,]. Only study procedures relevant to this validation are further elaborated here.

For both studies, participants completed in-laboratory and at-home portions; however, only the in-laboratory portions are considered for this work. Participants in the ICANGO study completed 2 in-laboratory visits spaced approximately 7-14 days apart, while those in the MAGIC study completed only 1 in-laboratory visit. During the in-laboratory visits, participants completed several walking tasks on an instrumented walkway (GAITRite, CIR Systems Inc) while also wearing a set of 6 wearable inertial sensors (OPAL v2, APDM Inc). Only the lumbar-mounted Opal sensor is considered for this work. Pressure data from the GAITRite system were collected at 100 Hz, while acceleration data from the Opal sensors were recorded at 128 Hz. A motion tracking system (SIMI, SIMI Motion Reality Systems) was also used during these in-laboratory walks; however, not all participants had analyzable data from this system. All 3 systems were synchronized via an electronic trigger to ensure precise time alignment.

For the in-laboratory walking tasks, participants completed 3 laps of the 20-foot GAITRite walkway at self-selected natural, slow, and fast walking speeds. Participants in the ICANGO study additionally completed a natural-speed walking task in which a carpet was placed over the GAITRite mat. This walk is considered only in the gait subblock analysis, as it was not included in either study.

Reference Algorithms

Data provided by the GAITRite mat were analyzed using the ProtoKinetics Movement Analysis Software (PKMAS) to obtain both gait subblock metrics and final gait metrics. Gait subblock metrics were those derived from the intermediate blocks of the gait algorithm, including but not limited to initial contact (IC) and final contact (FC) times. The final gait metrics of interest were limited to cadence, stride time, stride length, and gait speed, as these cover the majority of the temporal and spatial gait metric space.

Center of mass (CoM) displacement can often be used to simplify locomotion trajectories, as in walking, where modeling the CoM as an inverted pendulum aids in estimating spatial gait parameters such as stride length []. Reference data for CoM height change were provided by the SIMI system for those participants whose SIMI data were processed; these participants represented only a subset of those in the ICANGO study. Time points (ie, IC and FC events) used to calculate CoM height change were extracted from the PKMAS data.

The Comparison Gait Algorithm, Gait v2

In this work, the updated SKDH-gait, ie, gait v3, was compared with the last major version of the algorithm implemented in SKDH-gait, named gait v2. As a side note, the original GaitPy implementation is considered gait v1 [,], which is not included in the comparison, as it has already been sunset due to obsolescence.

Gait v2 uses continuous wavelet transforms (CWTs) and the inverted pendulum model to analyze the lumbar acceleration signal and generate gait metric estimates. Moving through the algorithm:

• Acceleration was first down-sampled to 50 Hz [].
• Acceleration frame orientation was corrected to better align with participant’s anatomical vertical, anterior-posterior (AP), and medial-lateral axes using a small-angle accelerometer tilt correction [].
• Following this orientation correction, a smoothed vertical acceleration signal was obtained by first integrating the vertical acceleration using the cumulative trapezoid rule and then applying the CWT with a Gaussian first-derivative wavelet [].
• Peaks in this smoothed vertical acceleration with a height above 50% of the SD of the signal were selected as first-pass estimates of the IC events (ie, heel strikes) [].
• The CWT was then applied to the smoothed vertical acceleration a second time to obtain an estimate of the smoothed jerk [].
• Peaks in this signal (again with a height >50% of the signal SD) were taken as first-pass estimates of the FC events (ie, toe-offs) [].

This first step for estimating IC and FC events will be referred to as the “m” method. The scales for the CWT were estimated from step frequency-CWT scale relations derived in previous work []. For this relation, an estimate of the step frequency was required, which was obtained using a pass of the algorithm at the scale from the original work [], defined by a frequency of 1.25 Hz (at 50 Hz, this scale is 8).

Using these first-pass estimates of ICs and FCs, steps are defined by applying the following quality control (QC) rules:

• Loading time (initial double support) should be less than C loading t max stride.
• Stance time should be less than half of t max stride plus the maximum initial double support from the first check.
• Strides should be less than the t max stride.

t_{max stride} is the maximum stride time and C_loading is a loading factor. t_{max stride} and C_loading were set to 2.25 seconds and 0.2, respectively, for the ICANGO study following previous work in gait “v1” []. For the MAGIC study, t_{max stride} was set to 1.75 seconds and the wavelet scale frequency was set to 1.35 Hz following previous internal work in pediatric populations.

Following the step construction from the ICs and FCs, the CoM height change between consecutive ICs was estimated by double-integrating the vertical acceleration. This CoM height change was then used in the inverted pendulum model to estimate step length [], as described below.

where S_step is the step length, l_leg is the leg length estimated from 0.53 × height [], and Δh is the CoM change in height. Step time was defined as the interval between consecutive ICs, and stride time as the sum of 2 consecutive step times. Stride length was defined as the sum of consecutive step lengths. Gait speed was calculated as stride length divided by stride time, and cadence was defined as the number of steps taken per minute.

Cadence = (60/t_Step)

where t_Step is the step time.

Comparison Contact Event Algorithms

In addition to the full comparison algorithm (gait v2), we implemented several previously published methods for estimating IC and FC events to provide additional benchmarks for our proposed method. These comparator methods are summarized in .

Proposed Gait Algorithm

Overview

Building on gait v2, we improved key components for estimating gait metrics from lumbar acceleration, resulting in the updated gait v3 algorithm. Although all data from the ICANGO and MAGIC studies were used in the final analysis of gait metrics, the subblock improvements were developed using a random ≈50% train-test split within each study. This yielded a training set of n=33 (14/19 from ICANGO/MAGIC) and a test set of n=47 (26/21 from ICANGO/MAGIC). Note that, in gait v3, acceleration data are not downsampled. presents the updated gait algorithm.

‎

Accelerometer Axis and Mean Step Frequency Estimation

The first major subblock in the proposed gait algorithm estimates (1) the accelerometer axis most closely aligned with the participant’s AP axis and (2) the mean step frequency for the entire walking bout. Identifying the AP axis is essential for applying the small-angle orientation correction and for subsequent contact event detection, which relies on the AP signal. Estimating the mean step frequency is also important, as it informs the dynamic QC thresholds based on walking speed and determines the appropriate CWT scale for smoothing the acceleration signal.

To estimate the AP axis, we first identified the vertical axis as the one with the largest absolute mean acceleration value. The acceleration signal was then filtered using a fourth-order band-pass Butterworth filter (0.25-7.5 Hz) to isolate walking-related components. After filtering, the AP axis was defined as the axis showing the highest cross-correlation with the vertical axis.

To estimate the mean step frequency, we computed CWTs of the AP-axis acceleration across 10 scales spanning 0.5-5.0 Hz using a Gaussian first-derivative wavelet. These 10 CWT coefficients were summed to produce a single signal representing the “CWT power” in that frequency range. A windowed (5 seconds, 50% overlap) autocovariance was then calculated, and for each window, the first peak was extracted. The median of these first peaks was defined as the mean step time, which was inverted to obtain the mean step frequency.

Foot Contact Event Detection

While the “m” method (The Comparison Gait Algorithm, Gait v2 []) for estimating contact events is generally reliable and yields good agreement for temporal gait metrics (ie, stride time), it is not the most accurate for determining the exact timing of these events []. shows the vertical and AP acceleration from a lumbar device, with several key points labeled.

‎

The issue arises from the smoothing of the vertical acceleration, which is very effective at locating point “V3” of the vertical acceleration []. However, this point does not correspond to the actual moment of IC, which is instead point “AP5” [,]. Therefore, we propose a new method that utilizes the AP-axis acceleration and dynamically adjusts the wavelet scale based on the mean step frequency for the walking bout.

First, we calculated the CWT scales based on the estimated mean step frequency per

where f_CWT is the Gaussian first derivative wavelet central frequency for the IC and FC detection, and is the estimated mean step frequency. The relations were obtained by performing a scale sweep on the training data and optimizing both the F₁-score and the root-mean-square error using a cost function []. We note that the generalizability of these relations to pathological cohorts should be further explored in future work. Although the step frequency-wavelet frequency relation only informs the CWT-based filtering based on walking speed, and therefore should not be affected by pathologies, this assumption should be confirmed.

To remove any very high-frequency noise, acceleration was filtered with a 20 Hz cutoff using a fourth-order Butterworth filter. Next, FC events were obtained by first integrating the filtered AP acceleration to compute the AP velocity, and then applying the CWT (Gaussian first-derivative wavelet) using the scale corresponding to the central frequency from the above equations. Peaks in the resulting CWT coefficients correspond to FC events. To minimize false positives, we applied 2 constraints to the peak detection: first, the prominence had to be greater than 0.6σ_Wn, and second, the peak distance had to be greater than , where σ_Wn is the SD of the CWT coefficients and is the estimated mean step time. These constraints were added after exploration of the training dataset during method development, based on visual inspection and experience, to ensure physiologically reasonable detections.

To obtain IC estimates, we first applied the CWT to the filtered AP acceleration using the scale corresponding to the estimated frequency, producing a smoothed AP jerk signal. Local minima and maxima were then identified using a prominence constraint of 0.6σ_Wn and a distance constraint . Finally, for each FC event, we searched backward to locate the corresponding IC event, defined as the point halfway between the preceding zero-crossing (positive to negative) and the preceding local minimum. This approach—rather than relying solely on local extrema—helps account for signal perturbations and flatter peaks that can occur during slow walking.

Step Creation and Physiological Thresholds

With the IC and FC event estimates obtained, we proceeded to stride creation and QC using physiological thresholds. The same basic rules as in gait v2, described above in “The Comparison Gait Algorithm, Gait v2” section, were applied.

These rules require defining 2 parameters: t_{max stride} and C_loading. While gait v2 uses static thresholds for these values, we explored dynamic functions based on the estimated mean step time. shows the relationship between mean step time and stride time from the reference data, which generally follows the expected 1:2 relation. From this plot, we observed that the original static threshold of 2.25 seconds did not adequately cover the full range of walking speeds, particularly during fast walking, where several strides could occur before reaching 2.25 seconds. Examining the SD of stride times, we found a maximum of approximately 0.5 25 seconds. Therefore, we set a new dynamic threshold based on the following:

This provides a threshold set at least 2 SDs above the expected mean stride time.

‎

illustrates the relationship between mean step time and the loading factor derived from the reference PKMAS gait data. The data show that the loading factor is not constant across walking speeds, decreasing noticeably as walking speed increases. To account for this, we implemented the following relation to calculate the loading factor:

‎

The coefficients for this relation were obtained using simple linear regression on the reference data. This approach provided a much-improved limit for initial double support times compared with the original threshold (), producing a more consistent gap between the mean initial double support time and the threshold across walking speeds. The blue line in represents the upper limit for allowable initial double support values and is therefore positioned above the mean to accommodate the variability of initial double support times within a walking bout.

Using these new threshold calculations, steps were created by identifying groups consisting of 1 IC and 2 FCs (opposite and same side) that satisfied the dynamic QC thresholds. We also ensured continuity, such that the same-side FC of one step served as the opposite-side FC for the subsequent step. Once steps were allocated in this manner, temporal gait metrics were computed for each step or for continuous steps, as appropriate.

Inverted Pendulum Model and Spatial Gait Metrics

Two inverted pendulum models are depicted in . Model 1 consists of a single inverted pendulum, whereas model 2 incorporates both a normal and an inverted pendulum. To improve spatial metric estimation in the proposed gait algorithm, we utilized the inverted pendulum model 2 []. This model can be described by the following equations [,]:

‎

One unknown remained on the right-hand side, l₂. To address this, we used the available temporal metrics to establish a relationship between the length ratio and the time ratio of these 2 gait phases, as follows:

where t_bip and t_sup represent the initial double support and single support durations for a step, respectively.

illustrates the experimental relationship described by the above equations, derived from reference PKMAS and SIMI data. From this training data, we obtained C=1.1 such that . Using this relationship, we estimated l₂ as follows:

l₂ = 1.1l_leg(t_bip/t_sup)

which enabled the calculation of the step length estimate using model 2. To minimize the variability in per-step estimations of the length-to-time ratio, we computed l₂ using the median t_bip and t_sup values across individual bouts of gait.

‎

Data and Statistical Analysis

Overview of the Analysis

The analyses were conducted in 2 parts: first, we computed values for each gait subblock described above and compared them with ground truth as well as the comparison algorithms; second, we computed final gait metrics for the proposed method (gait v3) and the comparison method (gait v2) and compared these with the reference.

For all analyses, when summarizing values across a walking task, the median value across all steps was used. When summarizing values across visits (for ICANGO), the average of the 2 visits was used. Intraclass correlation coefficients (ICCs) were used to evaluate agreement, with ICC≤0.4 indicating “poor,” <0.4-0.59 indicating “moderate,” 0.6-0.74 indicating “good,” and 0.75-1.0 indicating “excellent” agreement [].

Gait Subblock Performance Analysis

For the first part of the analysis, ICANGO and MAGIC participants were randomly divided into training and testing sets within each study (without accounting for age, gender, or other covariates) to develop the proposed method. This resulted in n=33 in the training set (14/19 from ICANGO/MAGIC), and n=47 in the test set (26/21 from ICANGO/MAGIC). Results for the gait subblocks are presented using the test set data. Additional results are provided in . Where appropriate, these results used outputs from the ground truth methods to isolate the effects of individual subblocks—for example, CoM height change estimation used reference method contact events as bounds for integrating acceleration.

For the comparison of mean step time estimation, we used ICC (2-way random effects, absolute agreement), Pearson correlation coefficients (Pearson ), and repeated measures correlation coefficients (RMCCs) to account for within-participant correlation (ie, multiple walking speed tasks per participant). Analysis of covariance was used to statistically adjust for interindividual variability [,].

To assess the performance of the gait subblock values, we considered both per-step values and median values across walking tasks. For IC and FC detection, performance was evaluated using sensitivity, precision, F₁-score, and mean absolute error (MAE) relative to the reference system. To determine true and false positives and negatives, each contact event in the reference system was matched to a corresponding event within a 0.3-second window (ie, 0.15 seconds on each side). Previous studies have used smaller windows (eg, 50 ms []), but such narrow windows would have excluded many events and artificially increased MAE. Events occurring outside the GAITRite mat area were not counted as false positives. MAE was computed only for contact events corresponding to a reference contact event. Performance metrics were assessed both visually and using linear mixed-effects regression (LMER) models to account for interactions between method, walking task, and age group, while controlling for within-participant correlation using a participant-specific random effect.

Gait Metric Performance Analysis

For the second part of the analysis, the full set of ICANGO and MAGIC participants was used, and gait metrics were compared between the 2 algorithm versions and the reference results. Gait v2 and gait v3 were computed using SKDH version 0.13.0. The SKDH code is available on GitHub []. The analysis focused on 4 common gait metrics—cadence, stride time, stride length, and gait speed—which together cover both temporal and spatial gait parameters.

To comprehensively evaluate the accuracy of gait metrics derived from gait v3 against both the reference system (PKMAS) and gait v2, we applied a hierarchical statistical framework to distinguish primary and secondary outcomes. Primary analyses were defined as agreement with reference values across walking speeds and were assessed using (1) Pearson correlation coefficients (r) to evaluate whether the algorithm preserved the relative ranking of individuals, and (2) ICC (2-way random effects, absolute agreement) to evaluate whether algorithm outputs were in absolute agreement with the reference. These measures directly capture the clinical interpretability of gait v3 relative to gold-standard assessments. Secondary analyses focused on evaluating whether gait v3 reduced systematic bias relative to gait v2. To this end, LMER models were used to compare PKMAS, gait v2, and gait v3 across walking speeds and age groups (>18, 12-17, and <12 years) within a comprehensive modeling framework. The 4 gait metrics (cadence, stride time, stride length, and gait speed) were modeled as outcomes, with algorithm, walk type, age group, and the interactions algorithm × walk type and algorithm × age group included as fixed effects. Participant-specific random intercepts were included to account for repeated measures. Pairwise contrasts of estimated means were calculated, with P values adjusted for multiple comparisons using the false discovery rate.

In the ICANGO study, where participants had 2 in-laboratory visits, we additionally assessed test-retest reliability using ICC (2-way random effects, absolute agreement), Pearson correlation coefficients, and RMCCs.

The following R packages (R Foundation) were used: rmcorr (version 0.6.0), lme4 (version 1.1-35.3), lmerTest (version 3.1.3), emmeans (version 1.10.1), rstatix (version 0.7.2), psych (version 2.4.3), and BlandAltmanLeh (version 0.3.1). R version 4.3.3 was used.

Ethics Considerations

Ethics approval for the MAGIC protocol was obtained from the Advarra Institutional Review Board (protocol number: Pro00047861). All participants provided informed consent. The ICANGO study was reviewed and approved by the Advarra Institutional Review Board (protocol number: Pro00043100).

Results

Mean Step Time Estimation

The agreement between the gait v3 mean step-time estimates and the reference method is presented in . The overall agreement was excellent, with all ICC values exceeding 0.91 and Pearson r values above 0.92. Repeated-measures analysis yielded similar findings, demonstrating a high RMCC of 0.98.

Initial and Final Contact Estimation

The accuracy of the proposed and comparison methods for IC/FC estimation relative to the reference methods—along with several additional literature-based methods (see )—is shown in A. The gait v3 method achieved among the highest F₁-scores, with median values above 0.9; the only comparable approach was the “m” method (gait v2). When comparing MAE across methods (B), gait v3 again performed among the best for IC estimation, with only the “g” method showing similar MAE values. The “pe” method approached this performance but generally produced higher errors. Notably, gait v3 showed an approximately 50% reduction in MAE compared with the “m” (gait v2) method. Full LMER model results comparing each algorithm with the gait v3 method are provided in .

‎

When comparing FC method F₁-scores (C), none of the literature-based methods approached the performance of either the “m” (gait v2) or gait v3 methods. Although both “m” and gait v3 exhibited some low F₁-score outliers, gait v3 had noticeably fewer outliers, and those that did occur were generally higher than the outliers observed for the “m” method. For MAE (D), the gait v3 method had the lowest values, with median MAE typically about 50% lower than any of the other methods.

Exploring the outliers reveals that most are not consistent across methods for either IC or FC estimation; rather, the outliers for different methods arise from different participant-walk tasks. Notably, for the gait v3 method, all F₁-score outliers remain above 0.81 for IC events and above 0.65 for FC events.

Gait Metrics

Using data from the ICANGO and MAGIC studies, we evaluated the performance of the gait v2 and v3 algorithms against reference PKMAS values. summarizes the algorithm and reference metrics, along with agreement statistics (Pearson correlation coefficients and ICCs) for each gait variable and walking task. Correlation coefficients were generally high and comparable between gait v2 and gait v3, with the lowest Pearson correlation of 0.73 for stride time during the slow walk using gait v2, and 0.82 for stride length during the slow walk using gait v3. By contrast, ICC values showed a consistent improvement with gait v3: while gait v2 exhibited a minimum ICC of 0.50 (gait speed, fast walk), gait v3’s lowest ICC was 0.81 (stride length, slow walk) across all metrics and walking tasks. The most notable improvements in ICC values were observed for the spatial metrics—stride length and gait speed. Highlighting a few specific instances, stride time during the slow walk showed a substantial increase, with Pearson r improving from 0.73 in v2 to 1.00 in v3, and ICC increasing from 0.70 to 1.00. Similarly, gait speed showed marked gains: during the fast walk, ICC improved from 0.50 to 0.84, and during the slow walk, from 0.71 to 0.90. These examples illustrate the consistent enhancement in agreement achieved with the gait v3 algorithm.

This improved agreement is illustrated in , where stride time results are generally consistent across methods and walking speeds. However, B shows that gait v2 aligns less closely with PKMAS results across all 3 gait speeds. Bland-Altman plots indicate that this difference reflects a mostly consistent bias, with gait v2 generally underestimating gait speed (see ).

‎

presents the pairwise contrasts of estimated means from the LMER models across walking tasks. When comparing gait v2 with PKMAS, most metrics showed significant differences (P<.007), except for cadence during the normal and slow walks, and stride time during the normal and fast walks. By contrast, gait v3 showed no significant differences from PKMAS for any metric across all walking tasks (the smallest P value was .14). Direct comparison of v2 and v3 revealed that the metrics where v2 differed from PKMAS also exhibited significant differences between v2 and v3 (P<.009). These significant differences in gait speed indicate that v2 was 0.064-0.199 m/s slower than v3, and 0.065-0.230 m/s slower than PKMAS.

shows a similar pattern, with significant differences between v2 and PKMAS for all age groups in stride length and gait speed. While v3 also showed some differences from PKMAS in these metrics, the discrepancies were smaller than those observed for v2. Direct comparison between v2 and v3 revealed that v2 had shorter stride lengths and slower gait speeds, with differences ranging from 0.134 to 0.156 m for stride length and 0.098 to 0.162 m/s for gait speed.

Test-Retest Reliability

We assessed the test-retest reliability of the reference, comparison, and proposed methods across walking speeds and gait metrics. Overall, the results in show excellent ICC agreement, with similar values across methods. Nearly all ICCs exceeded 0.75, except for gait v2 slow-walk cadence and gait v3 fast-walk stride time. The lower agreement for gait v3 stride time during the fast walk was influenced by an outlier, likely caused by accelerometer misalignment affecting the AP acceleration during the task. By contrast, gait v2 was not impacted because it relies on vertical acceleration. Even after correction, the misalignment resulted in fewer steps being properly detected. Other metrics maintained higher agreement, as they were either aggregate measures (eg, gait speed) or based on estimates from a larger number of steps (eg, cadence).

Discussion

In this work, we presented a step-by-step description of the updates made to the SKDH [] gait module, both within the algorithm subblocks and in the overall algorithm performance relative to a ground-truth GAITRite/PKMAS system. By applying the updated algorithm to healthy adult and pediatric populations in the ICANGO and MAGIC studies, and evaluating it across multiple comparison methods, these updates demonstrated significant improvements over the previous version of the gait module. This work introduces a generalizable, regulatory-grade framework for remote gait estimation using a single lumbar sensor—incorporating dynamic QC, robust axis-adaptive gait detection, and deployment-ready processing pipelines. These innovations move the field from algorithm development toward validated clinical deployment.

Estimation of the mean step time for a gait bout showed strong agreement, with higher Pearson correlation coefficients, ICCs, and RMCC. While per-stride step time is estimated as a final output of the gait algorithm, having an a priori estimate of the mean step time for the walking bout allows for dynamic adjustment of the gait event detection method as well as stride QC. This, in turn, minimizes the need for manual adjustment of algorithm parameters by the end user, which should support generalizability as the algorithm is deployed in patient populations. With these parameters—and those used in peak detection (eg, minimum peak prominence)—setting them based on the training set, whether visually or using ground-truth data, and then validating them in the test set provides confidence in their generalizability. However, future work might explore a sensitivity analysis of these parameters to assess their impact on the final results.

Accurate estimation of initial and final foot contact events is the critical foundational block for all subsequent gait metric calculations. We implemented a new approach for locating foot contacts, alongside implementations of several previously proposed methods for comparison [,,-,]. Our method demonstrated superior F₁-scores and MAE values compared with these methods across both IC and FC events. For IC events, the “g” (0.022 seconds) and “pe” (0.027 seconds) methods showed similar median MAE values; however, their F₁-score performance was substantially worse (0.93 and 0.68, respectively) compared with the proposed method (MAE 0.023 seconds; F₁-score of 0.98). FC results were even more favorable for the proposed method: only the “m” (gait v2) method achieved a comparable F₁-score (median 0.98 vs 1.0), and none of the other methods approached its MAE performance (gait v3 median 0.018 seconds). A review paper compiling results from various IC/FC detection algorithms across multiple populations reported MAE values for IC events ranging from 0.013 to 0.148 seconds in healthy participants, and from 0.025 to 0.186 seconds in individuals with neurologic diseases, Parkinson disease, and hemiplegia []. Both the “g” and “pe” methods were among the best-performing algorithms in that review, aligning with the performance of existing methods observed in this work.

These improvements to the gait processing subblocks would, however, have limited value if they did not enhance the overall estimation of gait metrics. Across the 4 gait metrics presented, we observe consistently better results for the proposed gait v3 algorithm. Correlation values generally remain stable or show improvements, such as for cadence during slow and fast walks. ICC values further highlight these gains—particularly for spatial metrics, where stride length and gait speed ICCs increase from 0.50-0.74 in gait v2 to 0.81-0.90 across walking speeds. We also observe improvements in agreement for temporal metrics during slow and fast walking (eg, stride time ICC during slow walking increases from 0.75 in v2 to 1.00 in v3). These strengthened agreements are likely the result of both dynamic gait event detection and QC using the a priori estimation of mean step time, which eliminates the need to manually adjust parameters across different walking tasks. This should enable more reliable gait monitoring in at-home environments, where walking characteristics may vary depending on context. The improvements also result in a reduction of bias in gait metric estimation. These gains were observed and confirmed in the mixed-effects models used to evaluate performance across walking speeds and age groups. Specifically, the differences between gait v3 and the reference PKMAS values are consistently reduced or eliminated compared with gait v2, indicating reduced estimation bias across walking speeds. Moreover, estimation bias in younger age groups was also reduced, as shown in . Although some differences in gait speed and stride length remain in gait v3, these discrepancies are smaller.

The importance of the statistically significant improvements and bias reduction achieved by gait v3 becomes clearer when interpreted against meaningful change thresholds or minimal clinically important differences, which vary across pathological cohorts. Specifically, across walking speeds and age groups, gait v3 reduced bias in gait speed estimates by approximately 0.10-0.20 m/s compared with v2. This range aligns with the European Medical Agency’s qualification of the 95th percentile of gait speed, where changes of 0.10-0.20 m/s are considered clinically meaningful in Duchenne muscular dystrophy []. Similarly, in Parkinson disease, clinically important differences in gait speed have been reported up to ~0.20 m/s, depending on the assessment method []. The fact that the reductions in bias observed with gait v3 fall within these clinically meaningful ranges indicates that the updated algorithm enhances the ability to detect meaningful changes that gait v2 may have obscured. This improvement in accuracy is particularly important in clinical trial contexts, where misestimation at the level of clinically meaningful differences can inflate sample size requirements or mask therapeutic effects, and in real-world monitoring, where longitudinal changes must be confidently attributed to underlying clinical status rather than algorithmic error.

Beyond technical improvements, gait v3 also reflects a distinct design philosophy compared with conventional gait pipelines. Many existing approaches rely on multisensor fusion, subject-specific calibration, or data-driven modeling to optimize accuracy under controlled conditions—methods that can limit scalability and reduce compliance in real-world or decentralized settings. By contrast, gait v3 adopts a physically motivated, single-site lumbar design with embedded biomechanical constraints and dynamic QC, enabling calibration-free, interpretable, and robust performance across heterogeneous users and environments. This shift from calibration-dependent precision to calibration-free robustness emphasizes scalability, compliance, and longitudinal continuity—key requirements for home-based monitoring and decentralized clinical trials. By delivering reliable gait metrics from a single self-applied sensor, gait v3 reduces participant and site burden while maintaining regulatory-grade accuracy. Thus, its contribution extends beyond performance gains: it illustrates how biomechanically grounded, dynamically self-correcting algorithms can reconcile analytical rigor with real-world feasibility, advancing the broader adoption of digital gait end points in clinical research. In summary, gait v3 is well-suited to the emerging landscape of patient-centric, real-world digital health applications and decentralized clinical trials.

The most significant recent effort in gait algorithm design and validation was undertaken by the Mobilise-D collaboration, which recently published the technical validation of their gait platform in healthy adults as well as individuals with Parkinson disease, multiple sclerosis, chronic obstructive pulmonary disease, congestive heart failure, and proximal femur fracture. When evaluating multiple subblock algorithms, the best-performing method for IC detection demonstrated a sensitivity of 0.80 and absolute errors of 0.06 seconds in healthy participants, with similar performance across pathological cohorts []. Although the gait v2 algorithm implemented in SKDH has not been published in patient populations, the Mobilise-D results showed a wide range of ICC values for IC estimation (0.68-0.98) and step length (0.28-0.70) when validating the originally published algorithms on which gait v2 was based []. We report lower MAE (under 0.05 seconds) and higher F₁-scores (over 0.95), although the values reported by Mobilise-D were obtained during simulated daily activities rather than prescripted tasks, which would likely contribute to the higher performance observed here. Following their gait subblock assessment and algorithm exploration, the Mobilise-D consortium also published full gait speed results based on the subblocks that performed best in their testing []. Within their full validation protocol—including simulated daily activities—algorithm performance during straight walking showed errors of –0.03 m/s (limits of agreement –0.18 to 0.11), with an absolute error of 0.07 m/s (limits of agreement 0.01 to 0.13) in healthy adults. This is very similar to our reported error of 0.024 m/s (95% CI –0.03 to 0.08) for normal straight-walk gait speed. Across all reported in-laboratory tasks, the Mobilise-D algorithm yielded an ICC of 0.74 (95% CI 0.66-0.81) for gait speed, slightly lower than our observed 0.82-0.90 across walking speeds, although the inclusion of simulated daily activities likely accounts for this difference. Overall, we report results that are comparable to, or better than, those from the Mobilise-D project, suggesting that SKDH may be a suitable alternative—pending further validation in pathological cohorts and in settings involving more simulated daily activities. This is particularly relevant for the dynamic thresholds used for QC checks. Although these thresholds were designed to be broad enough to encompass a wide range of physiological parameters, their underlying relationships may not always hold in pathological gait. Nevertheless, we maintain that dynamic, data-driven thresholds based on observed walking behavior are likely to outperform the original static thresholds in gait v2 when applied to pathological gait.

Assessing test-retest reliability across the 2 in-laboratory visits in the ICANGO study, ICC values were consistent for both the reference method and gait v3. This provides evidence to support its use as a reliable clinical outcome assessment and as a digital end point in clinical trials [].

There are several limitations to this work. First, because the study population consisted solely of healthy individuals, we cannot infer the generalizability of the algorithm updates to patient populations. Although the inclusion of children and adolescents provides some indication of broader applicability, it remains insufficient to establish generalizability to clinical cohorts. Second, validation was restricted to prescripted in-laboratory walking tasks, even though multiple walking speeds were included. To address these limitations, future work will evaluate the algorithm’s validity in patient cohorts with distinct gait signatures, assess performance during simulated daily living, examine the impact of sensor misplacement, and compare results with other clinical outcome measures.

(Bloomberg) — US stocks extended losses as a selloff in the year’s biggest artificial intelligence winners dragged global gauges back from the brink of record highs. Longer-dated bond yields climbed.

A disappointing sales outlook from chipmaker Broadcom Inc. weighed on rivals and further fueled investor anxiety over AI wagers initially sparked by Oracle Corp. The AI bellwether’s stock drop started Thursday following a forecast for rising outlays and a longer timeline to a revenue payoff. The slump deepened on a report of delays to some of Oracle’s data center projects Friday. Shares of companies tied to the power infrastructure also slid.

The Nasdaq 100 fell as much as 2.3% while Broadcom Inc. sank 11%. The S&P 500 fell more than 1% after the index notched a record close in the previous session. The Dow Jones Industrial Average pulled back from record highs.

“The AI bubble is deflating but not popping,” said Louis Navellier, chief investment officer at Navellier & Associates. Friday’s profit-taking after stock benchmarks hit new highs is not unexpected, he added. “Heightened concerns on OpenAI agreements may make further gains challenging.”

Investors also had to contend with mixed messages from Federal Reserve officials after they delivered their third-straight interest rate cut this week.

Yields on Treasury 30-year bonds rose 6 basis points to a three-month high after Federal Reserve Bank of Cleveland President Beth Hammack said she would prefer interest rates to be slightly more restrictive to keep putting pressure on inflation, which is still running too high.

Kansas City Fed President Jeff Schmid also pointed to consumer pricing pressures. Schmid said he dissented against the central bank’s decision this week to lower interest rates because inflation remains too high and the economy continues to show momentum.

“Judging from the calendar, this could be a day Wall Street takes a breather from its recent feverish pace,” said Joe Mazzola, head trading and derivatives strategist at Charles Schwab. “Major earnings and data are scarce, the weekend beckons, and investors increasingly look ahead to next Tuesday’s jobs report.”

Diversification across geographies and themes is becoming a key consideration. After technology heavyweights drove equity gains for much of the year, concerns about stretched valuations and vast capital outlays have prompted investors to look for opportunities elsewhere.

“Given the set-up in markets, diversification is now the price worth paying to keep you fully invested in equities,” wrote Goldman Sachs’s Mark Wilson. He adds that there are compelling investment stories including Korea, Japan, China or the broader emerging markets.

Meanwhile, strategists at Goldman Sachs Group Inc. expect stocks to notch fresh records next year, citing resilient economic growth and broader adoption of artificial intelligence to support corporate earnings.

The team led by Ben Snider reaffirmed its target for the S&P 500 to reach around 7,600 points in 2026, implying gains of about 10% from current levels. Other forecasters and asset managers share the upbeat view, with strategists at firms including Morgan Stanley, Deutsche Bank AG and RBC Capital Markets LLC also calling for US stocks to rise more than 10%.

Market forecasters are broadly positive on Europe as well, with not a single one of the 17 strategists surveyed by Bloomberg expecting a major decline. Four strategists, including those at UBS Group AG and Deutsche Bank AG, project gains of nearly 13% from Wednesday’s close.

Some are eyeing gains on an even shorter horizon, betting on further advances before 2025 ends as investors rotate into stocks that have so far remained in tech’s shadow.

“Everyone is convincing themselves that there will be a Christmas rally, so it looks like there will be one, and to be honest, there’s no negative catalyst visible until the end of the year,” said Karen Georges, a fund manager at Ecofi Investissements in Paris. “Investors are keen to buy this year’s laggards, it’s a good time to diversify your portfolio at the moment.”

Bloomberg’s index of the dollar clinged to a gain after a two-day losing streak and was on track for a third weekly loss. In commodities, gold erased gains while silver pulled back from an all-time high.

What Bloomberg Strategists Say…

“Oracle is an outlier among the major AI stocks as it has run down its cash flow and taken on debt to finance capital spending. If the model is running into trouble already because of difficulties in getting data centers built, that’s bad news for all the companies selling shovels into this gold rush, as well as potentially the cloud giants themselves, which have valuations partly predicated on huge and growing order backlogs.”

—Sebastian Boyd, Macro Strategist, Markets Live

For the full analysis, click here.

Corporate News:

A group of influential Swiss lawmakers proposed watering down the capital demands that the country wants to impose on UBS Group AG, sending shares to a 17-year high. Broadcom Inc., a chip company vying with Nvidia Corp. for AI computing revenue, slumped after its sales outlook for the red-hot market failed to meet investors’ lofty expectations. Lululemon Athletica Inc. shares rallied after the pricey yoga-wear maker boosted its full-year outlook and announced that its chief executive officer would step down after a period of sluggish growth. SoftBank Group Corp. is studying potential acquisitions including data center operator Switch Inc., people with knowledge of the matter said, underscoring billionaire founder Masayoshi Son’s growing ambitions to ride an AI-fueled boom in digital infrastructure. Some of the main moves in markets:

Stocks

The S&P 500 fell 0.9% as of 1:03 p.m. New York time The Nasdaq 100 fell 1.6% The Dow Jones Industrial Average fell 0.3% The MSCI World Index fell 0.7% The Russell 2000 Index fell 1% Philadelphia Stock Exchange Semiconductor Index fell 4% Currencies

The Bloomberg Dollar Spot Index was little changed The euro was little changed at $1.1747 The British pound was little changed at $1.3375 The Japanese yen fell 0.1% to 155.79 per dollar Cryptocurrencies

Bitcoin fell 2.9% to $90,222.32 Ether fell 5.4% to $3,075.98 Bonds

The yield on 10-year Treasuries advanced four basis points to 4.19% Germany’s 10-year yield advanced one basis point to 2.86% Britain’s 10-year yield advanced three basis points to 4.52% The yield on 30-year Treasuries advanced six basis points to 4.86% Commodities

West Texas Intermediate crude rose 0.1% to $57.66 a barrel Spot gold rose 0.4% to $4,296.55 an ounce Silver Future Mar26 fell 4.1% to $61.92 This story was produced with the assistance of Bloomberg Automation.

–With assistance from Andre Janse van Vuuren, Levin Stamm, Julien Ponthus, Jan-Patrick Barnert, Michael Msika, Sagarika Jaisinghani and Ira Iosebashvili.

©2025 Bloomberg L.P.

The future of work is changing fast. Future Focus cuts through the noise with three trends each week that matter most to HR and business leaders. When everything else is in flux, stay focused with Future Focus.

Longer Leash On Life: Inside The Dog Longevity Startup (Forbes) 

What to Know: A biotech startup developing canine longevity drugs has raised $135 million, is pursuing FDA conditional approval, and is running the largest animal clinical trial with 1,300 dogs. Its first pill aims to mimic caloric restriction in senior dogs, with additional candidates targeting large-breed growth hormone pathways — potentially hitting the market as early as 2026. 

Where to Focus: Pet-related benefits are increasingly part of total rewards packages. Breakthroughs in pet health and longevity — even in early biotech spaces — could further signal rising employee expectations around pet insurance, caregiving leave, and wellness supports.

Research: 2025 Employee Benefits Survey

Why Young People are Struggling to Communicate (Time) 

What to Know: Educators report accelerating declines in teens’ writing and speaking skills, driven by reduced face-to-face interaction and heavy reliance on social media and generative AI. Research cited suggests chatbot use can reduce mental effort and creativity, further weakening original thinking and real-time communication. 

Where to Focus: The next wave of early-career hires will have sharper access to tools but shallower practice in core human skills. Leaders should expect higher demand for structured communication training, more deliberate in-person collaboration, and clearer AI-use norms — because communication, not code, remains the backbone of decision-making, client trust, and leadership pipelines.

Related article: A Guide to Leading an Effective Multi-Generational Workforce

Vibe coding has come to Washington (Fast Company) 

What to Know: Figma’s AI prototyping tool, Make, is now available to Figma for Government users. It aims to compress prototype cycles from weeks to hours as agencies race to improve services across more than 10,000 federal websites. With FedRAMP authorization and growing government adoption, AI-assisted “vibe coding” is set to influence how public-sector digital services are designed and delivered. 

Where to Focus: When the public sector standardizes on AI design workflows, expectations for speed, accessibility, and usability may rise across the economy. Vendors and enterprises that serve regulated industries could experience shorter procurement and delivery windows, prioritize compliance-ready toolchains, and upskill product and design teams to work credibly with AI because the new baseline for service quality will be set in public view.

Be the first to know what’s next. 

Get Future Focus delivered weekly through Tomorrowist — your shortcut to smarter decisions in a changing world of work.

Several products are being recalled due to potential safety risks, according to the U.S. Consumer Product Safety Commission (CPSC).

Companies like Sanven Technology are voluntarily pulling certain items off the shelves because they might be dangerous.

Here’s what you need to know to keep your family safe:

Sanven – VEVOR

Around 2,840 VEVOR garment steamers are being pulled because they may leak or spray hot water, posing a serious burn hazard.

The CPSC said the water tank lid can also come loose, exposing users to hot water.

Six incidents have occurred, resulting in three burn injuries. One caused a permanent scar, the CPSC said.

These were sold online through Amazon.com, HomeDepot.com and Vevor.com from December 2023 to September 2025.

You can get a full refund by emailing Sanven Technology with contact information and a photo of the steamer as directed.

You can find the recall information here.

Around 11,650 VEVOR ice crushers are also being recalled because they may experience a thermal event and catch fire.

Two thermal events have been reported, with one resulting in a fire, the CPSC said. No injuries were reported.

The CPSC said these were sold online through Amazon.com, Walmart.com, Wayfair.com, Lowes.com and others from July 2021 to July 2025.

If you have one, you can email recalling@vevor.com with pictures of the affected product for a full refund.

You can find the recall information here.

YCXXKJ

Around 8,960 of the company’s baby bath seats are being pulled because they may be unstable and tip over, posing a drowning hazard.

No incidents or injuries have been reported, according to the CPSC.

These were sold on Amazon.com from May 2024 to October 2025.

You can email bentalkhere@outlook.com for a full refund. Make sure to follow the company’s directions when sending photos.

You can find the recall information here.

Read more Consumer news on KSAT:

Australia’s New Vehicle Efficiency Standard (NVES) began on January 1, 2025, setting CO₂ emission targets for light-duty vehicles through 2029. While the NVES targets are based on the New European Driving Cycle (NEDC), Australia is also adopting Euro 6d-equivalent exhaust emissions standards that require vehicles to be tested using either the Worldwide harmonized Light vehicle Test Procedure (WLTP) or the U.S. 2-cycle tests.  

To avoid requiring the auto industry to test vehicles under multiple protocols, the Australian Department of Infrastructure, Transport, Regional Development, Communications and the Arts engaged ICCT in December 2024 to develop conversion methods that translate the WLTP and U.S. 2-cycle CO₂ values into NEDC-equivalent values for compliance determination with NVES.   

ICCT developed detailed conversion algorithms covering the various Australian fleets, including petrol and diesel passenger cars, light commercial vehicles, and plug-in hybrids. This test cycle conversion protocol was adopted into Australia’s legislation as an option for the auto industry to use for type approval CO₂ conversions instead of physical testing on multiple test cycles, if preferred. 

The report can be viewed and downloaded from the Australian government website. 

Pakistan on Friday signed a memorandum of understanding (MoU) with global crypto exchange Binance to explore tokenisation of sovereign and real-world assets of up to $2 billion, marking one of the country’s most significant steps toward adopting blockchain-based financial infrastructure, PTI reported.The MoU was signed at the Finance Division by Finance Minister Muhammad Aurangzeb and Binance CEO Richard Teng, the Ministry of Finance said in a statement. The agreement aims to assess the feasibility of digitising Pakistan’s assets — including government bonds, treasury bills, commodity reserves and other federally owned holdings — through blockchain platforms.According to the ministry, the proposed collaboration seeks to “enhance liquidity, transparency and international market accessibility,” with any move subject to Pakistani laws, policies and regulatory approvals. The MoU is non-binding, sets no exclusivity, and does not amount to a procurement commitment. Definitive agreements, if pursued, will be negotiated within six months.Officials said the arrangement could allow Binance and its affiliates to provide technical expertise, advisory support, training and capacity building to help Pakistan evaluate compliant blockchain infrastructure that can attract global investors while ensuring full sovereign control.Calling the MoU a major signal of the government’s reform direction, Aurangzeb said, “This is a very strong message — not only for Pakistan, but for the entire world. What we have signed today reflects a long-term partnership… The next step for us is execution, and we are fully committed to delivering results with speed and quality.”Advisor to the Pakistan Crypto Council Changpeng Zhao, who was present at the signing, termed the development “a landmark” for Pakistan’s financial future. “This is a great signal for the global blockchain industry and for Pakistan… now we can move towards full deployment and execution,” he said, expressing confidence that the collaboration would have “positive and lasting outcomes for the economy.”The ministry added that the initiative marks a key milestone in aligning Pakistan’s financial ecosystem with global best practices at a time when tokenisation is gaining traction worldwide.Separately, the Pakistan Virtual Assets Regulatory Authority (PVARA) said it has issued No Objection Certificates to Binance and HTX, launching a phased, FATF-aligned route toward full licensing. “Strong governance, AML and CFT compliance remain central as Pakistan builds a trusted digital asset ecosystem,” PVARA said on X.The authority said Pakistan has taken a “decisive step” toward a regulated digital-asset framework by entering formal arrangements with Binance and HTX.

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

The writer is an FT contributing editor

In early December, the US state of New Hampshire announced what it called the “world’s first bitcoin-backed municipal bond”. Munis — bond issues from American states and cities — are a $4.3tn market. They’re a minor part of a $47.8tn market for bonds in the US, dominated by Treasuries and corporate issues, but munis confer a special blessing. Any interest earned from a muni is exempt from federal income taxes, and some munis even allow the lender to escape state income taxes.

This tax break, $183bn in total over the next four years and long contested in US politics, allows local governments to borrow at lower rates. If a corporation can convince a local government to borrow on its behalf it can enjoy those lower rates, too, sheltered in what’s essentially a federal subsidy for private business.

These “private activity bonds” are a common strategy to encourage companies to invest in infrastructure like hospitals or utilities. States also often have an even broader understanding of what infrastructure is, though — and which companies could use cheaper loans.

The New Hampshire bitcoin muni has cracked open the door to a new idea: companies that hold bitcoin can do something for local economic development. This could lure bitcoin treasury companies or more bitcoin miners up north, at a time when the state legislature is considering a ban on any local restrictions on server farms.

The new muni from New Hampshire stretches the definition of what economic development is. Digital assets are interesting. It’s not clear how they create growth.

Sovereigns have always wanted to spend money they don’t yet have, and have had an interest in creating markets for their debt. In the early American republic, states chartered commercial banks as an economic development strategy. In return for a literal licence to print dollars, the bank would be obliged to dig a canal, for example. That approach fell out of favour with the widespread failure of chartered banks in 1837, and states increasingly turned to the federal practice of issuing bonds.

Congress reconsidered the tax exemption on state bonds several times in the 20th century, particularly when raising revenue for the first and second world wars, but the principle that states could raise money unmolested by Washington endured.

During the Great Depression, states had started offering this shelter as “industrial development bonds,” issues on behalf of local champions. This practice boomed in the middle of the century, growing from $6mn in 1956 to over $500mn just a decade later. In 1968 and again in 1986, Congress finally curbed this practice, more clearly defining what a private activity bond is and limiting issues by state.

In 2022, the latest year for which the IRS offers data, states issued $119bn in private activity bonds. Most of these went to build airports, low-income housing and hospitals, and to support non-profits. New Hampshire’s Business Finance Authority, which issues the state’s private activity bonds, still has about $85mn in capacity under these limits, which means it has room to be creative, just as states feel the pressure to find more creative ways to borrow.

“It’s not getting any easier to be a state anytime soon,” says Matt Fabian of Municipal Market Analytics, a research firm. States are more frequently running deficits and muni issues are growing, he explains, as states take on more obligations such as losses from climate change. This leaves less capacity for economic development.

Normally, private activity bonds are guaranteed by future revenues — from the company and ultimately the state. In New Hampshire’s bond, a company would post its bitcoin as collateral, to be held in trust. The state issues the bonds, essentially converting the private company’s bitcoin into a cash loan.

Bitcoin has lost 14 per cent of its value over the past six months; the head of the state’s business finance authority has said that the private company would have to put up bitcoin worth 150 per cent of the bond issue.

“It seems interesting,” says Daniel Garrett, assistant professor of finance at Wharton, “but it doesn’t make any immediate sense to me as to why it would be necessary.” Garrett, whose research looks into the development returns from municipal bonds, says he can’t figure out the economic or public value the state gets out of it.

New Hampshire, like Arizona and Texas, has approved a crypto reserve fund, which would allow it to hold bitcoin in the state’s rainy day fund, a vehicle some US states maintain for unexpected expenses. Together with the plan for a new bond issue, the reserve fund positions the state as friendly to crypto.

Here again, it’s hard to understand what the state is getting in return for making an alternative currency liquid. It could just as easily hold euro-denominated assets in its rainy day fund or convert euro holdings to cash through a muni issue. Munis can be a creative way to get private companies to do things, but until we can figure out what the private companies will be doing, the question for New Hampshire’s new bond issue is the same as for most innovations around crypto: “Cool. But why?”