Development and comparative validation of multiple models for cognitive frailty in older adults residing in nursing homes

Objectives This study aims to develop an optimal predictive model for cognitive frailty (CF) in older adults residing in nursing homes, thereby providing a scientific basis for staff to assess CF risk and implement preventive interventions. Methods This study recruited 500 older adults from four nursing homes in Hangzhou, Zhejiang Province, between December 2024 and March 2025 as the modeling cohort. Additionally, we enrolled 112 older adults from another nursing home in Hangzhou from March to April 2025 as the external validation cohort. With 19 variables, we applied k-nearest neighbors (KNN), support vector machine (SVM), logistic regression (LR), random forest (RF), and extreme gradient boosting (XGBoost) algorithms to forecast CF. The predictive performance was assessed through multiple evaluation approaches, including ROC curve evaluation, calibration curve assessment, decision curve analysis, and various classification metrics such as accuracy, precision, recall, Brier score, and the F1-score (with β = 1). Furthermore, Shapley additive explanations (SHAP) value analysis was performed for the optimal model. Results Among 500 older adults in nursing homes, 132 (26.4%) exhibited CF. Essential features included the activities of daily living (ADL), frequency of intellectual activities, and age, among others. Five models using different algorithms were developed. The SVM model demonstrated the best predictive performance, with an AUC of 0.932 on the test data. External validation confirmed its accuracy (AUC = 0.751). Conclusion Machine learning models, particularly SVM, can effectively predict CF risk in older adults residing in nursing homes. Care facility staff can utilize personal information to assess older adults and identify high-risk individuals for CF at an early stage, providing crucial support for timely interventions and quality of life enhancement.