An interpretable machine learning model predicts frailty risk in middle-aged and older adults with gastrointestinal disease: a longitudinal study

Abstract

Frailty is a significant health concern in the aging global population, particularly among middle-aged and older adults with gastrointestinal disease (GID). Early detection of individuals at increased risk is critical for implementing timely preventive and therapeutic interventions. This study aimed to develop and validate an interpretable machine learning (ML) model to assess frailty risk in this population. To overcome the “black box” nature of conventional ML models, we integrated Shapley Additive exPlanations (SHAP), which helps identify key predictors of frailty and improve the interpretability of the model’s decision-making process. This study analyzed data from the 2013-2015 survey waves of the China Health and Retirement Longitudinal Study (CHARLS). To identify the most predictive variables for frailty, we employed a dual-method approach combining the Boruta algorithm and Least Absolute Shrinkage and Selection Operator (LASSO) regression. We applied ten different ML algorithms to develop prediction models: Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Classifier (GBC), Light Gradient Boosting Machine (LightGBM), K-Nearest Neighbors (KNN), Decision Tree (DT), Multilayer Perceptron (MLP), Naive Bayes (NB), and Adaptive Boosting (AdaBoost). The area under the receiver operating characteristic (ROC) curve (AUC) served as the primary performance metric. Additional metrics, including sensitivity, specificity, precision, and the F1-score, were used to comprehensively evaluate model accuracy. Calibration curves were generated to assess the consistency between predicted probabilities and observed risk, and the Brier score was used as a quantitative measure of calibration accuracy. Decision curve analysis (DCA) was performed to evaluate the clinical net benefit of each model. To understand the impact of individual predictors on the output, the SHAP method was used to provide transparent insights into each feature’s contribution to the estimated frailty risk. A total of 1,404 participants met the eligibility criteria for this study, of whom 444 (31.62%) were classified as frail. Using the Boruta algorithm and LASSO regression, we identified 10 key predictors of frailty. Among all the ML models tested, the LR model showed the best overall performance, achieving an AUC of 0.759 (95% CI: 0.711–0.806). Shapley Additive exPlanations (SHAP) analysis further revealed the top five predictors of frailty in this population: depression, grip strength, education level, the total number of chronic diseases, and self-rated health. This study introduces an interpretable ML model that effectively detects frailty risk among middle-aged and older adults with GID. The model demonstrates strong predictive accuracy and transparency, supporting its potential as a clinical decision-support tool pending further external validation and real-world deployment. Such proactive measures could improve patient care and promote better long-term health outcomes in this population.