Abstract / Summary
Abstract Background Frailty is a significant geriatric syndrome that adversely affects postoperative recovery in older patients undergoing hip arthroplasty. This study aimed to identify distinct frailty trajectories in older adults after hip arthroplasty, develop and validate a machine learning-based prediction model for these trajectories, and identify key influencing factors using explainable artificial intelligence methods. Methods A retrospective cohort of 347 patients aged ≥65 years undergoing hip arthroplasty was enrolled between January 23, 2021 and August 12, 2024. Frailty was assessed using the Frailty Index (FI) at multiple time points. Latent growth mixture modeling (LGMM) was employed to identify distinct frailty trajectories. Eight machine learning algorithms were compared to predict trajectory classification, with model performance evaluated by Area Under the Curve (AUC), accuracy, sensitivity, and F1-score. SHapley Additive exPlanations (SHAP) were used to interpret the best-performing model. Results Three distinct frailty trajectories were identified: Worsening Group (17.29%, n = 60), Gradual Recovery Group (40.06%, n = 139), and Rapid Recovery Group (42.65%, n = 148). For subsequent binary classification, the two recovery groups (GRG and RRG) were merged into a single Improvement Group, and a random forest model was developed to distinguish improvement from worsening. The random forest model demonstrated the best performance among the evaluated models (internal validation AUC = 0.819,external validation AUC = 0.821). SHAP analysis revealed that Oxford Hip Score, Mini-Mental State Examination (MMSE), Connor-Davidson Resilience Scale (CD-RISC) score, social support, and albumin levels were key protective factors, while advanced age, high CCI score, depressive symptoms, and elevated fasting blood glucose were significant risk factors. Conclusion Three distinct postoperative frailty trajectories were identified following hip arthroplasty in older adults. A random forest prediction model incorporating multidimensional baseline predictors demonstrated favorable discriminatory performance (AUC = 0.819), with SHAP analysis identifying key modifiable predictors including cognitive function, psychological resilience, social support, and nutritional status. These findings confirm the study hypothesis and provide a foundation for preoperative risk stratification, though validation in larger multicenter cohorts is required before clinical translation.