Machine learning and SHAP-based prediction of depression among residents in Gansu Province

Abstract Background Depression is a leading cause of disability worldwide, yet community-based screening and early identification remain inadequate. Conventional statistical approaches assume linear relationships and are limited in capturing the complex interactions between factors. Methods We evaluated five supervised classifiers via nested 10-fold cross-validation on data from the 2024–2025 Gansu Provincial Mental Health Literacy and Common Mental Disorders Survey, and interpreted the final model using SHapley Additive exPlanations (SHAP).Restricted cubic splines were used to model nonlinear dose-response relationships, and additive and multiplicative interactions were quantified using RERI, AP, synergy index, and 3D probability surface plots. Results XGBoost achieved satisfactory discriminative performance (AUROC = 0.884; AUPRC = 0.328; sensitivity = 0.690). SHAP analysis identified insomnia as the most influential predictor, followed by mental health literacy(MHL) attitude and age. Restricted cubic spline analysis demonstrated significant nonlinear trends for insomnia, MHL attitude, and age. Insomnia and poor MHL attitude exhibited synergistic interaction, while insomnia and age showed antagonistic interaction. Conclusion Integrating XGBoost with SHAP explainability yields both accurate prediction and mechanistic transparency. The findings challenge conventional linear assumptions and support a dual-target prevention strategy combining insomnia management with MHL attitude modification, alongside age-stratified screening in community settings.