Machine learning-based classification of frailty among adults with mild cognitive impairment in China: a cross-sectional study

Introduction Frailty is associated with adverse health outcomes among adults with mild cognitive impairment (MCI). Accurate identification of prevalent frailty in this population is clinically important. This study developed and compared machine-learning (ML) models for frailty classification using multidimensional, routinely available features. Methods We included 3,124 adults aged ≥ 45 years with MCI from the 2015 China Health and Retirement Longitudinal Study. Participants were divided into training and test sets at a 7:3 ratio. Missing-data imputation, feature selection, and model development were conducted using the training set only. Candidate features were selected using LASSO regression, and eight ML models were compared. Classification performance was evaluated using the area under the receiver operating characteristic curve (AUC), precision–recall area under the curve (PR-AUC), sensitivity, specificity, F1-score, Brier score, calibration analysis, and decision curve analysis. Robustness was assessed using bootstrap resampling of the test set and repeated cross-validation within the training set. Logistic regression (LR) coefficients and odds ratios, together with SHapley Additive exPlanations (SHAP) analysis of XGBoost, were used for complementary model interpretation. Results Frailty was identified in 13.0% of participants (406/3,124). LASSO retained six features: activities of daily living (ADL), instrumental activities of daily living (IADL), self-reported health (SRH), age, pain, and sleep disturbance. LR achieved the highest test-set AUC (0.887) and PR-AUC (0.537), whereas XGBoost showed similar discrimination, with an AUC of 0.885, a PR-AUC of 0.522, and the highest F1-score (0.535). Bootstrap and repeated cross-validation analyses showed no clear discrimination advantage of either LR or XGBoost. SHAP analysis identified ADL, IADL, SRH, age, pain, and sleep disturbance as contributors to the XGBoost classification output. Conclusions LR and XGBoost showed similar discrimination in classifying prevalent frailty among adults with MCI, with different strengths across performance metrics. Routinely available functional and patient-reported features may support multidimensional frailty assessment. Because the study was cross-sectional, the models should not be interpreted as predicting future frailty onset. Prospective external validation is required before clinical application.