Development and external validation of an interpretable machine learning model for depressive symptom screening in Chinese adults aged 50 years or older with diabetes

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background: Depressive symptoms are common among people with diabetes and may impair self-management and quality of life. Objective: This study aimed to develop and externally validate an interpretable machine-learning model for depressive symptom screening in Chinese adults aged 50 years or older with diabetes and translate it into an online assessment tool. Methods: Participants with diabetes from Round 4 of the China Health and Retirement Longitudinal Study (n = 2,187) were randomly divided into training and internal validation/model-selection sets at a 7:3 ratio. An independent cohort from Xijing Hospital (n = 695) was reserved for external validation. Depressive symptoms were defined as a Center for Epidemiologic Studies Depression Scale-10 (CESD-10) score of 10 or higher. Least absolute shrinkage and selection operator regression was used for feature selection; nine machine-learning algorithms were compared in the development data, and SHapley Additive exPlanations analysis interpreted the final Extreme Gradient Boosting (XGBoost) model. Results: Sixteen predictors were retained. XGBoost achieved areas under the curve of 0.822, 0.832, and 0.812 in the training, internal validation, and external validation sets, respectively. Major contributors included sleep duration, self-perceived health status, cognitive function, activities of daily living, instrumental activities of daily living, and life satisfaction. An online assessment tool was developed. Conclusions: The XGBoost model showed consistent discrimination but non-ideal calibration and may support concurrent preliminary classification and triage of depressive symptoms in Chinese adults aged 50 years or older with diabetes. The calculator is not diagnostic, should not replace CESD-10 administration or professional assessment when clinically indicated, and requires recalibration and prospective multicenter validation before broader clinical use.