Explainable AI for Population Mental Well-being Surveillance Using Community Health Survey Data

Population mental well-being surveillance increasingly relies on machine learning methods. However, most existing studies focus on predictive accuracy using black-box models, offering limited transparency and producing findings that are difficult to translate into public health action. This study proposes an explainable AI framework for identifying high-confidence, human-readable patterns of mental well-being outcomes using the Canadian Community Health Survey 2019-2020 data. We analysed 108252 anonymised survey records with approximately 50 attributes capturing demographics, chronic conditions, lifestyle behaviours, and psychosocial factors. After removing non-informative attributes and consolidating sparse response categories, we trained interpretable C4.5 decision tree models for four outcomes: pain status, stress level, work stress, and the Health Utility Index. The proposed approach achieved strong predictive performance across outcomes, with accuracies of 87.0\% (pain), 82.1\% (stress), 95.76\% (work stress), and 82.3\% (functional health). To enable actionable surveillance insights, all decision paths were automatically extracted as rules and annotated with support and confidence, revealing consistent co-occurrence patterns linking functional limitations and musculoskeletal conditions with pain, and highlighting associations between life satisfaction, sense of belonging, age strata, and elevated stress even among respondents reporting good mental health. Overall, the findings demonstrate that large-scale national health surveys can be effectively leveraged for explainable mental well-being surveillance, delivering interpretable evidence to support population risk stratification, early-warning monitoring, and policy-relevant public health interventions.