Development of explainable machine learning-based prediction models of depression with mild and severe insomnia in...

The risk of depression is significantly elevated in middle-aged and older adults with insomnia; however, the pathways between mild and severe insomnia are different. Most existing prediction tools treat insomnia as a single construct, ignoring its heterogeneity. It is crucial to develop an interpretable, severity-specific modeling framework to quantify depression risk, identify shared determinants, and guide targeted prevention efforts.

We analyzed 2433 participants (≥45 years) from the China Health and Retirement Longitudinal Study. Insomnia was classified as mild (5-6 h) or severe (<5 h). Depression was measured with the 10-item CES-D. After LASSO selection of demographic, clinical, and lifestyle predictors, nine machine-learning algorithms were trained and compared; interpretability was provided by SHapley Additive exPlanations (SHAP).

Random Forest performed best (AUC 0.7784 for mild insomnia, 0.7554 for severe insomnia), with good calibration and decision-curve utility. Older age, female gender, poor self-rated health, and multimorbidity consistently elevated risk, whereas life satisfaction, functional independence, cognitive reserve, and education were protective. Both models are deployed in a Streamlit web app that delivers individual risk scores and SHAP-based explanations in real time.

Our study introduces parallel, explainable random forest models that stratify depression risk by insomnia severity, revealing universal vulnerabilities alongside severity-specific drivers. Implemented in a web-based tool with SHAP visualization, the framework supports rapid, stratified screening and tailored counseling in community and primary-care settings with limited specialist resources. By linking interpretable AI with severity-aware insights, this work advances precision prevention of depression in sleep-disturbed middle-aged and older populations.