Machine learning-based risk prediction models for depressive symptoms in Chinese community-dwelling individuals with multimorbidity: A longitudinal study

Background: The increasing prevalence of physical-mental multimorbidity poses a significant challenge to healthcare systems. This study aimed to develop and validate machine learning (ML) algorithms to predict depressive symptoms among individuals with multimorbidity in China. Methods: Data were extracted from the China Health and Retirement Longitudinal Study. Depressive symptoms were assessed using the Center for Epidemiological Studies Depression (CES-D) scale. We employed four ML algorithms to construct prediction models, and calculated feature importance to identify key predictors of depressive symptoms. Results: Depressive symptoms were observed in 593/1816 (32.65 %) of the participants with multimorbidity at 8-year follow-up. The LASSO regression identified nine significant predictors. The areas under the curve for models were 0.724 for logistic regression (LR), 0.638 for random forest (RF), and 0.675 for eXtreme gradient boosting (XGBoost). LR and XGBoost demonstrated good overall performances, with LR performing slightly better. Calibration curves indicated high model accuracy, and decision curve analysis confirmed their clinical utility. The models performed well within the cardiometabolic pattern. SHapley Additive exPlanations identified five features: baseline CES-D, cognitive function, grip strength, chair-rising time, and economic development region. A nomogram was developed to visualize the predictive factors for depressive symptoms. Limitations: External validation is required to substantiate the models' generalizability. Conclusion: Through this study, we offer a user-friendly model for identifying depressive symptoms in community-dwelling individuals with multimorbidity. Effective prediction of depressive symptoms enables early screening of at-risk populations, facilitating timely intervention to mitigate occurrence risk and reduce healthcare costs.