Machine Learning-Based Prediction of Post-PKP Frailty: A Retrospective

Introduction

Frailty, a clinical syndrome characterized by diminished physiological reserves and multisystem dysfunction, compromises the capacity to withstand minor stressors, thereby predisposing individuals to adverse health outcomes.^1,2 Although prevalent across all age groups, frailty demonstrates age-dependent epidemiology, with incidence rates positively correlating with advancing age.³ Global demographic projections estimate the population aged ≥65 years will reach 2 billion by 2050, suggesting an impending escalation in frailty burden.⁴ Substantial evidence links frailty to critical adverse outcomes including falls, increased hospitalization rates, malignancies, and mortality, which collectively deteriorate elderly health and impose substantial strain on healthcare systems.^5–9 Notably, frailty represents a dynamic and potentially reversible state except during terminal decline phases, underscoring the imperative to identify modifiable risk factors for targeted prevention and mitigation strategies.¹⁰

Osteoporotic vertebral compression fractures (OVCFs), a prevalent geriatric condition, have emerged as a critical public health concern due to their substantial morbidity burden in elderly populations.¹¹ Substantial evidence identifies frailty as a significant predictor of osteoporotic fractures in older adults.¹² Furthermore, OVCFs exhibit an accumulative effect on both frailty incidence and progression, establishing a bidirectional causal relationship between these two clinical entities.¹³ However, the differential impacts of post-OVCF intervention strategies on frailty development remain underexplored in current clinical research.

Since its inception in the 1990s, percutaneous kyphoplasty (PKP) has been refined through decades of clinical practice and is now established as a gold-standard intervention for OVCFs.^14,15 PKP achieves rapid analgesia, stabilizes fractured vertebral bodies, and promotes early mobilization, thereby reducing immobilization-related complications such as pneumonia and thromboembolism.^16–18 Nevertheless, procedure-associated adverse events—including refractures, cement leakage, neural compromise, and adjacent-segment vertebral fractures—pose non-negligible risks.^19–22 Currently, whether perioperative PKP affect frailty development remains unknown.

Machine learning (ML) has been widely implemented in disease prognosis prediction, with recent advances showing superior performance in forecasting spinal surgery outcomes compared to traditional statistical models.²³ This study applied machine learning to identify key perioperative factors influencing post-PKP frailty development, providing actionable insights for optimizing surgical approach selection and technique modification to reduce frailty incidence.

Methods

Patients and Study Design

The study conducted a retrospective analysis of a dataset obtained from our hospital, spanning from April 2013 to April 2023. The external validation dataset covers April 2017 to April 2022 from Wenzhou Hospital of Integrated Traditional Chinese and Western Medicine. All patients in this study signed an informed consent form prior to undergoing procedure. And ethics approval was obtained from Ethics Committee in Clinical Research (ECCR) of the First Affiliated Hospital of Wenzhou Medical University (KU2025-R062) and Wenzhou Hospital of Integrated Traditional Chinese and Western Medicine (2025-L134; for external validation). The study was conducted according to the Declaration of Helsinki.

Inclusion criteria were as follows: (1) OVCFs resulting from low-energy trauma (including sprains, heavy lifting, or falls) with MRI-confirmed acute vertebral fractures; (2) T score of bone mineral density (BMD) < −2.5; (3) Patients underwent PKP surgical intervention.

Exclusion criteria were as follows: (1) history of spinal surgery; (2) presence of frailty at admission; (3) pathological vertebral fractures secondary to malignancies or infections; (4) fractures with spinal cord compression accompanied by neurological deficits (eg, numbness and/or muscle weakness); (5) incomplete medical records; (6) cognitive impairment precluding independent communication; (7) incomplete follow-up data.

Baseline Data and Primary Outcome

In the study, population characteristics [age, sex, weight, height, body mass index (BMI), smoking, drinking, hypertension, diabetes, and hyperlipidemia], perioperative imaging features (severe osteoporosis, the number of fractured vertebrae, segmental classification of spine fractures, Genant classification, T2 hyperintensity in soft tissue, restoration of anterior vertebral height, and kyphosis, preoperative and immediate postoperative Visual Analogue Scale (VAS) scores, anesthetic features [American Society of Anesthesiologists Physical Status Classification System (ASA classification), type of anesthesia, and anesthesia duration], and surgical and perioperative information (surgical time, intraoperative blood loss, volume of bone cement used, bone cement extravasation, duration of postoperative recumbency, and duration of hospitalization). In addition, residual low back pain (LBP) was evaluated according to following criteria during the one-year follow-up period: defined as persistent pain occurring in the original pain site postoperatively, with a VAS score of ≥4.

The 2-year interval sufficiently meets both PKP prognostic assessment criteria and longitudinal frailty monitoring requirements.^24–27 So the primary outcome was postoperative frailty after PKP at two-years follow-up. Frailty was diagnosed according to the Fried criteria.¹

Radiographic Characteristics

1. Severe osteoporosisDual-energy X-ray absorptiometry (DXA)-derived T-scores exceeding −3.5.²⁸
2. Genant classification²⁹
3. T2 hyperintensity in soft tissueHyperintensity in the dorsal soft tissues at or below the fracture level was observed on preoperative T2-weighted MRI sequences.
4. Restoration of anterior vertebral heightThe anterior vertebral body height of the fractured vertebra was measured on sagittal radiographs both preoperatively and postoperatively, with the difference quantified as the restoration.
5. KyphosisKyphosis angle was quantified on sagittal radiographs by measuring the angle between the superior endplate of the adjacent superior vertebra and the inferior endplate of the adjacent inferior vertebra. Kyphosis was defined as a postoperative-to-3-month angular change >10°.

Feature Selection

The guidelines for developing ML models were followed for the present study.³⁰ The patient cohort that underwent PKP was divided into training and testing subsets with a ratio of 4:1. Consequently, we obtained a Dataset 1 and a Dataset 2, with 3679 and 920 patients respectively.

To alleviate the potential bias caused by our limited sample size, the least absolute shrinkage and selection operator (Lasso) regression was employed to select features.²³ Features were selected based on the λ.1-s.e. (standard error) criterion.³¹ And λ serves as the penalty parameter, determining the degree to which the function is shrunk. Lasso regression coefficient of each variable was shown in Supplementary Table 1.

Model Training and Testing

The model was established using five ML algorithms below: XGBoost (Extreme Gradient Boost), LightGBM (Light Gradient Boosting Machine), AdaBoost (Adaptive Boosting), GBDT (Gradient Boosting Decision Tree), and SVM (Support Vector Machine) with the RBF (radial basic function) kernel. Based on the area under the receiver operating characteristics curve (AUC-ROC), the optimal hyperparameters were optimized by grid search to improve the predictive performance of our models.^23,32 And the hyperparameters for each model were shown in Supplementary Table 2.

The testing set was used to evaluate the predictive power of each model. These assessments were performed using various metrics, including AUC-ROC, accuracy, sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV). In brief, discrimination was assessed using AUC-ROC for binary classification.³³ Calibration of the ML models was evaluated by calibration plots. The Brier score was recognized as a measurement of the models’ discrimination and calibration.³⁴ Additionally, the net benefit of each model was analyzed using decision curve analysis.³⁵ Ultimately, utilizing the ML model that demonstrated the best predictive performance, we ranked and visualized the contribution of each predictor to the model’s predictive performance using Shapley Additive Explanations (SHAP) value analysis.³⁶

Statistical Analysis

In the study, mean ± SD was used to represent continuous variables and frequency [proportion (%)] was used to represent categorical variables. Intergroup comparisons were employed using Student’s unpaired t-test and Chi-Square tests. Comparisons with values of P < 0.05 were considered statistically significant. Lasso regression was conducted using R programming language (4.2.3). And AdaBoost, GBDT, and SVM were trained using Python (scikit-learn = 1.1.3). XGBoost, LightGBM, and SHAP value analyses were respectively trained using Python (xgboost = 2.0.1), Python (lightgbm = 3.2.1), and Python (shap = 0.43.0).

Result

Patient Data

This study enrolled 4599 patients from a single-center cohort and 561 from an external validation cohort who underwent PKP. Comprehensive demographic profiles and clinical characteristics of both cohorts are summarized in Table 1.

Baseline Predictive Performance

Based on the λ.1-s.e. criterion, seven features were selected using Lasso regression (Figure 1 and Supplementary Table 1) for the subsequent model developing. The resulting features encompassed residual LBP, severe osteoporosis, Genant classification, T2 hyperintensity in soft tissue, the number of fractured vertebrae, and the duration of postoperative recumbency.

Figure 1 (A) The coefficient profiles of the 29 features were analyzed by Lasso regression. (B) Based on the λ.1-s.e. criterion, there were seven features selected using Lasso regression. The two dotted vertical lines were drawn at the optimal scores by minimum criterion (λ.min) and λ.1-s.e. criterion, respectively. At the λ.1-s.e. criterion, the selected features included residual LBP, severe osteoporosis, Genant classification, T2 hyperintensity in soft tissue, the number of fractured vertebrae, and the duration of postoperative recumbency. Abbreviations: Lasso, the least absolute shrinkage and selection operator; s.e., standard error; LBP, low back pain.

The default hyperparameters for each model were shown in Supplementary Table 2. The comparative analysis of the baseline predictive performance of the ML models in the validation set (10-fold cross-validation) was shown in Supplementary Figure 1 and Supplementary Table 3. As shown in Figure 2 and Table 2, the XGBoost model demonstrated the superior baseline predictive performance in testing sets, achieving an AUC-ROC [95% confidence interval (CI)] of 0.946 (0.930–0.963), an accuracy of 0.882, a sensitivity of 0.885, a specificity of 0.881, a PPV of 0.647, a NPV of 0.969, and a Brier score of 0.072. Furthermore, the calibration curve of the XGBoost model was fitting the perfect situation (Supplementary Figure 2). And the decision curves indicated that XGBoost, GBDT, and SVM models performed well in clinical usefulness (Supplementary Figure 3).

Table 2 Comparison of Brier Score, AUC-ROC, Accuracy, Sensitivity, Specificity, PPV, NPV of the Baseline Prediction Performance for ML Models in Testing Sets

Figure 2 The baseline predictive performance for ML models in the testing dataset. (A) The ROC curves for ML models. (B) The AUCs for ML models. Abbreviations: ML, machine learning; ROC, receiver operating characteristics; AUC, the area under curve; XGBoost, Extreme Gradient Boost; LightGBM, Light Gradient Boosting Machine; AdaBoost, Adaptive Boosting; GBDT, Gradient Boosting Decision Tree; SVM, Support Vector Machine.

The Performance of ML Models with the Optimal Hyperparameters

The hyperparameters of each model were optimized by grid search. And the optimized hyperparameters for each model were shown in Supplementary Table 2. The comparative analysis of the ML predictive performance after hyperparameter optimization in the validation set (10-fold cross-validation) was shown in Supplementary Figure 4 and Supplementary Table 4. Interestingly, after hyperparameter optimizing, the XGBoost model still demonstrated the best predictive performance in the testing set, with an AUC-ROC (95% CI) of 0.950 (0.934–0.965), an accuracy of 0.884, a sensitivity of 0.890, a specificity of 0.882, a PPV of 0.651, a NPV of 0.970, and a Brier score of 0.066 (Figure 3 and Table 3). And the calibration curve of the hyperparameter-optimized XGBoost model was very close to the ideal calibrated curve (Supplementary Figure 5). As shown in Supplementary Figure 6, XGBoost, LightGBM, GBDT, and SVM models possessed excellent clinical benefit.

Table 3 Comparison of Brier Score, AUC-ROC, Accuracy, Sensitivity, Specificity, PPV, NPV of Each ML Model After Optimizing Hyperparameters Using Grid Search in Testing Sets

Figure 3 The predictive efficiency of ML models with the optimized hyperparameters using the grid search. (A) The ROC curves for hyperparameter-optimizing ML models. (B) The AUCs for hyperparameter-optimizing ML models. Abbreviations: ML, machine learning; ROC, receiver operating characteristics; AUC, the area under curve; XGBoost, Extreme Gradient Boost; LightGBM, Light Gradient Boosting Machine; AdaBoost, Adaptive Boosting; GBDT, Gradient Boosting Decision Tree; SVM, Support Vector Machine.

The performance of the hyperparameter-optimized XGBoost model, trained as described, remained stable in the external test set (AUC: 0.845, 95% CI (0.805–0.884)) (Figure 4 and Table 4). Moreover, in the external test set, the calibration curve of the hyperparameter-optimized XGBoost model also showed strong alignment with the ideal curve (Supplementary Figure 7) and the model still possessed excellent clinical benefit (Supplementary Figure 8).

Table 4 The Performance of Hyperparameter-Optimized XGBoost Model for Predicting Post-PKP Frailty in the External Test Set

Figure 4 The ROC curve of the hyperparameter-optimized XGBoost model for predicting post-PKP frailty in the external test set. Abbreviations: ROC, receiver operating characteristics; AUC, the area under curve; XGBoost, Extreme Gradient Boost; CI, confidence interval.

In addition, based on the hyperparameter-optimized XGBoost model, we identified the importance ranking of each Lasso regression-selected feature using the SHAP value analysis. The results were shown in descending order as follows: residual LBP, Genant classification, the duration of postoperative recumbency, severe osteoporosis, T2 hyperintensity in soft tissue, and the number of fractured vertebrae (Figure 5).

Figure 5 A SHAP analysis of the XGBoost model with the optimized hyperparameters. (A) A visualization for SHAP values of the Lasso-regression-selected features in each sample. The dots in red represented the higher SHAP values, while the blue ones represented the lower values. (B) The selected features were listed in descending order by the average absolute values of SHAP. Abbreviations: SHAP, Shapley Additive Explanations; LBP, low back pain.

Discussion

In recent years, linear regression and nomograms have been extensively utilized for clinical prognostic prediction and have demonstrated favorable performance in predicting postoperative outcomes following PKP.^22,37 Machine learning, which employs extensive datasets to optimize algorithmic performance, demonstrates superior reliability, objectivity, and reproducibility when processing large-scale clinical data³⁸ Emerging evidence suggests that machine learning algorithms outperform conventional statistical methods in predicting clinical outcomes following PKP.³⁹ However, previous studies failed to implement variable selection prior to model development, potentially leading to overfitting and multicollinearity issues. LASSO regression effectively addresses these limitations. In this study, we employed LASSO regression with the λ.1-standard error (λ.1-s.e.) criterion for feature selection, followed by predictive model construction using the identified variables. Our findings identify six significant predictors of postoperative frailty: residual LBP, Genant classification, duration of postoperative recumbency, severe osteoporosis, T2 hyperintensity in paraspinal soft tissues, and number of fractured vertebrae. This model facilitates personalized postoperative rehabilitation protocols tailored to individual patients, providing evidence-based guidance to interprofessional rehabilitation teams to mitigate frailty risk in PKP populations.

Notably, residual LBP was identified as the most significant predictor of frailty following PKP, consistent with previous studies demonstrating the association between LBP and frailty.⁴⁰ Residual LBP persistently impairs postoperative recovery and disrupts sleep patterns, with prolonged pain and insomnia contributing to anxiety and depression.⁴¹ Furthermore, LBP alters gait characteristics, reduces movement coordination, and induces kinesiophobia. Pavel et al demonstrated that LBP may compromise diaphragmatic function, thereby affecting trunk stability.⁴² These factors collectively reduce physical activity levels and increase sedentary behavior, ultimately elevating frailty risk.⁴⁰ Progressive frailty inevitably diminishes health-related quality of life (HRQoL).⁴³ We emphasize that enhanced postoperative pain management and prevention of residual LBP could reduce frailty incidence and provide long-term clinical benefits for patients.

The Genant classification and number of fractured vertebrae, as preoperative indicators, reflect the severity of vertebral fractures. Since its introduction in 1993,²⁹ the Genant classification has been widely utilized for assessing OVCFs. Higher grades indicate more severe vertebral deformities, characterized by greater buckling of the vertebral cortex or endplates.⁴⁴ Multilevel vertebral fractures suggest extensive trauma to the spinal system. Severe Genant classifications and a higher number of fractured vertebrae correlate with poorer surgical outcomes, increased incidence of kyphosis, and significant associations with frailty.⁴⁵ Previous studies have also demonstrated the relationship between severe osteoporotic fractures, multilevel vertebral fractures, and frailty, which aligns with our findings.⁴⁶ Notably, this study found no correlation between kyphosis progression and frailty, potentially because single-level kyphosis progression does not reflect overall spinal deformity.

T2 hyperintensity in soft tissue typically indicates muscle edema, hemorrhage, or inflammatory exudates, reflecting functional impairment of the lumbar musculature. Although PKP allows early postoperative ambulation, muscle injury may restrict patient mobility and prolong recovery. Additionally, prolonged bed rest during hospitalization increases frailty risk, underscoring the importance of early mobilization to prevent disuse atrophy.

Substantial evidence confirms that osteoporosis exacerbates frailty syndrome in older adults through systemic chronic inflammation, nutritional alterations, and endocrine system dysregulation. Chronic inflammation acts as a pivotal determinant of frailty, exerting direct effects while also mediating indirect impacts via intermediate mechanisms.^47,48 Consequently, timely anti-osteoporosis therapy is imperative for patients with severe osteoporosis following PKP.⁴⁹ Beyond pharmacological interventions, balanced physical activity and dietary modifications constitute essential components for preventing or reversing frailty. For patients with established frailty, treatment strategies should be tailored based on specific frailty domains.⁵⁰

While our study demonstrates the potential of ML in predicting postoperative frailty, several limitations should be acknowledged. First, the study was retrospective in nature, and the data were collected from a single institution, which may limit the generalizability of the findings. We cannot exclude potential spectrum bias and selection bias in the tertiary medical center cohort, while unmeasured perioperative rehabilitation variables may potentially confound frailty outcomes. Future studies should validate our findings in larger, multicenter, and prospective cohorts. Second, the study focused on a specific PKP for OVCFs. The predictive performance of ML models may vary for other surgical interventions or patient populations. Third, while we identified key predictors of postoperative frailty, the underlying mechanisms remain unclear. Moreover, the internal dataset exhibited class imbalance that may introduce potential biases in ML model training and optimization—a limitation to be systematically addressed in our subsequent research. At last, this study employed a binary classification of frailty status (frail/non-frail) without stratifying the pre-frail subgroup for comparative analysis. Future research should explore the biological pathways linking these predictors to frailty development.

Conclusion

Our findings indicated that the application of ML algorithms, particularly the hyperparameter-optimized XGBoost model, can effectively predict postoperative frailty in patients underwent PKP. Furthermore, the key predictors were identified through SHAP value analysis. Standardized anti-osteoporosis treatment, prevention of residual postoperative low back pain, and reduction of duration of postoperative recumbency are critical factors in avoiding the occurrence of frailty following PKP.