併存うつ病および不安症状（CDAS）における要因効果の違い：個別化メンタルヘルス促進のための機械学習アプローチ

Depression and anxiety comorbid presents greater severity than isolated symptoms. Traditional analyses often fail to capture complex interactions among multidimensional risk and protective factors.

To develop a predictive model distinguishing young Chinese adults with comorbid depressive and anxiety symptoms (CDAS) from asymptomatic individuals using machine learning within Antonovsky's salutogenic framework, identifying protective factors as Generalized Resistance Resources (GRRs) and heterogeneous risk subtypes for precision-targeted interventions.

Data from 2951 Chinese young adults (aged 19-35) from the PBICR2021 database were analyzed. Fifteen features spanning psychological, physiological, social, and environmental dimensions were selected through Lasso Regularization and the Boruta algorithm. Six machine learning models were trained and evaluated. SHAP analysis interpreted key predictive factors, with bidirectional hierarchical clustering identifying risk subtypes. Decision curve analysis validated clinical utility.

Random Forest achieved optimal performance (AUC = 0.905, PR-AUC = 0.703), with good calibration (Brier Score = 0.097, ECE = 0.088). SHAP revealed differential association patterns: solo exercise (swimming, sports equipment), avoidant emotional regulation, obesogenic behaviors and stress were positive linked to CDAS among Chinese youth, while self-efficacy, family health, social support and healthcare access (as GRRs) were protective. Hierarchical clustering analysis of SHAP values identified three CDAS subtypes with distinct factor combination patterns. Decision curve analysis validated clinical utility within 2-67 % risk thresholds.

Protective and risk factors showed opposite associations between comorbid versus asymptomatic groups among Chinese young adults. Protective factors operationalize Antonovsky's GRRs, while three heterogeneous subtypes support precision-targeted preventions. The 15-feature model demonstrated clinical utility within 2-67 % risk thresholds for scalable community screening.