Risk prediction of depression- and anxiety-related psychological distress in a large multistage survey sample of...

To develop an interpretable ensemble machine-learning model to support risk stratification of elevated depression- and anxiety-related psychological distress in a large multistage survey sample of Chinese adults, and to examine nonlinear associations with key psychosocial factors.

We analysed an unweighted analytical sample of 10,294 adults from the China Social Mentality Survey. Psychological distress was measured with the Kessler-10 scale, and 13 psychosocial indicators were included as candidate predictors. The sample was split 7:3 into training and hold-out test sets. Six binary classifiers were trained and combined using a TOPSIS-weighted classifier fusion (TCF) ensemble. SHAP with LOWESS was used to examine feature importance and dose-response patterns.

The TCF model showed the most balanced overall performance across metrics. SHAP analyses showed that loneliness and dehumanization experiences were the strongest risk features, whereas positive psychosocial resources were generally associated with lower predicted risk. SHAP-LOWESS curves indicated nonlinear dose-response relationships: predicted risk rose when key adverse psychosocial features exceeded specific thresholds, whereas several positive resources showed protective patterns within certain ranges.

In a large unweighted multistage survey sample of Chinese adults, the TCF ensemble model provided an interpretable and balanced prediction framework for depression- and anxiety-related psychological distress and highlighted loneliness, dehumanization, and psychosocial resources as key correlates of elevated distress risk. These nonlinear patterns and thresholds may inform future risk assessment and prioritisation of psychosocial assessment, but require confirmation in longitudinal, externally validated, and design-weighted cohorts.