Development and external validation of interpretable machine learning models for predicting first fall risk in knee osteoarthritis patients: evidence from the CHARLS cohort and a hospital-based cohort in China.

Background

Knee osteoarthritis (KOA) is a prevalent degenerative joint disease that causes pain, reduced mobility, and an increased risk of falls in older adults. Predicting fall risk in this population remains difficult because of the complex interplay of multiple risk factors.

Methods

Data from 1097 patients with KOA in the CHARLS cohort were retrospectively analyzed using 28 demographic, health, and behavioral variables. Ten machine learning (ML) algorithms were evaluated, and the Gradient Boosting Machine (GBM) was selected as the final model based on overall balanced performance. Model interpretability was assessed using SHapley Additive exPlanations (SHAP), which identified the main predictors. External validation was performed in an independent hospital-based cohort to assess generalizability.

Results

The GBM model achieved the most balanced and stable performance (AUC = 0.836), demonstrating good discriminative ability. Sensory impairments (hearing and vision problems) and depressive symptoms were the most influential predictors of falls. SHAP analysis clarified the contribution of these features to the model's predictions. In the external cohort, the locked GBM model maintained favorable discrimination and calibration, confirming robustness and generalizability.

Conclusions

An interpretable ML model was developed to predict fall risk in KOA patients. The model effectively identified key sensory and mental health factors and demonstrated reliable performance across internal and external validations, supporting its potential to assist clinicians in early identification and prevention of falls among older adults with KOA.