Predicting frailty trajectories in older adults after hip surgery using interpretable machine learning (Preprint)

BACKGROUND Postoperative frailty is highly prevalent among older adults undergoing hip surgery and is closely linked to poor clinical outcomes. Despite growing interest in understanding its progression, the temporal patterns of frailty remain underexplored. Moreover, there is a lack of validated models that can predict frailty trajectories and stratify patients by risk in the early postoperative period. OBJECTIVE This study aimed to identify distinct frailty trajectories within six months following hip surgery in older adults and to explore their associated predictors. An interpretable machine learning model was developed and validated for individualized risk prediction and was implemented as a clinically accessible web-based calculator. METHODS This prospective longitudinal observational study was conducted among older adults undergoing hip surgery at a tertiary hospital in China. Frailty assessments were performed preoperatively and at 1, 3, and 6 months postoperatively. A total of 209 participants who completed the six-month follow-up were included in the analysis. Frailty was assessed using the Frailty Index (FI), and group-based trajectory modeling (GBTM) was applied to identify distinct frailty progression patterns. Predictive variables were selected using least absolute shrinkage and selection operator (LASSO) regression. An interpretable extreme gradient boosting (XGBoost) model was developed using a 60:40 training–test data split. Model performance was evaluated in terms of discrimination, calibration, and clinical utility. Interpretability was assessed using SHapley Additive exPlanations (SHAP) at both the global and individual levels. RESULTS Three distinct frailty trajectories were identified: low-fluctuation frailty, high-improvement frailty, and high-deterioration frailty. Twelve predictors grounded in the Health Ecology Model were selected, spanning individual characteristics, interpersonal networks, and the living environment. The XGBoost model demonstrated excellent discrimination, with a micro-average area under the receiver operating characteristic curve (AUC) of 0.98 (95% CI: 0.96–0.99) in the training set and 0.93 (95% CI: 0.90–0.96) in the test set. Calibration was acceptable, with a weighted Brier score of 0.0852. Decision curve analysis showed favorable clinical utility across a range of threshold probabilities. A web-based risk calculator was developed to facilitate personalized frailty trajectory prediction. CONCLUSIONS The XGBoost model demonstrated strong predictive performance and interpretability, enabling early identification of older patients at risk for adverse frailty trajectories following hip surgery. This tool may support targeted interventions and improve perioperative care in geriatric populations.