Assessment of disability in the older adults using electronic health record–based data: a machine learning approach

Background Global population aging is accelerating, and China faces a significant challenge with “aging before affluent.” The disabled older adults population in China is projected to reach 58 million by 2050. Early identification of disability risk is essential for optimizing healthcare resource allocation and improving long-term care systems. Objective This study aimed to develop and validate machine learning models using comprehensive electronic medical record (EMR) data to predict four distinct levels of disability in older adults inpatients. Methods Data from 523 patients (age ≥ 60) were retrospectively collected from two tertiary hospitals in Northeast China. Disability was categorized into four levels (A–D) based on the Barthel Index (BI). Feature selection was performed using LASSO regression with 10-fold cross-validation. Six algorithms—Logistic Regression (LOG), Random Forest (RF), Gradient Boosting Machine (GBM), XGBoost, AdaBoost, and Support Vector Machine (SVM)—were evaluated. SHapley Additive exPlanations (SHAP) was employed to interpret model decisions. Results Seventeen critical predictors were identified, including age, appendicular skeletal muscle mass index (ASMI), phase angle, and various inflammatory markers. The LOG and XGBoost models demonstrated the best performance (AUC = 0.83; Accuracy = 0.91). Ten-fold cross-validation confirmed stable model. SHAP analysis indicated that neurological comorbidities, muscle mass indicators, and nutritional-inflammatory status were the primary drivers of disability risk. Conclusion EMR-based machine learning models, particularly XGBoost, provide a robust and interpretable tool for early disability risk stratification, supporting clinical decision-making.