Beyond chronological age: a decision tree model integrating nutritional and renal function indicators for risk prediction in pre-frail older adults with hepatolithiasis.

Objectives: This study develops a decision tree model to predict risks and identify key factors in pre-frail older adults hepatolithiasis patients and performance of models. Methods: The study included 451 pre-frail older adults. Two balanced sample groups totaling 294 cases were obtained via Propensity Score Matching (PSM). Risk prediction models for hepatolithiasis were constructed based on the full sample and the two groups via applying the Classification and Regression Tree algorithm. Models' performance was evaluated by 10-fold cross-validation, ROC, and AUC, and models' performance and key factors for hepatolithiasis risk were compared and analyzed. Results: A total of 451 pre-frail older adults were included. PSM (1:1 ratio) was conducted to reduce confounding factors, with 147 matched pairs. In the full-sample decision tree model, age (>66.85 years) was identified as the primary split node, followed sequentially by strongly positive urine protein, marital status (married, widowed), height, and educational level. The model achieved a classification accuracy of 72.4%, with an area under the curve (AUC) of 0.795 (95% CI: 0.744-0.845). Following PSM, two subgroup decision tree models were developed. The Group 1 model identified age (>66.45 years), body mass index (>21.40 kg/m Conclusion: Decision tree risk prediction models for hepatolithiasis were developed in pre-frail older adults. Age emerged as the dominant risk factor across all models, with additional predictors involving nutritional, metabolic, and renal function indicators. Propensity score-matched subgroup models demonstrated superior discriminative performance compared to the full-sample model. These findings indicate that decision tree modeling, especially when combined with PSM, provides an interpretable and effective approach for identifying high-risk individuals in this vulnerable population. Nevertheless, the potential for overfitting underscores the necessity of external validation using larger, independent cohorts to ensure model generalizability. Future efforts may integrate ensemble learning algorithms to enhance model stability and stratify populations based on advanced age combined with metabolic or hepatic dysfunction profiles, thereby enabling more targeted prevention and intervention strategies for hepatolithiasis.