Machine Learning Models Predict Depression Risk Across Differing Air Quality Environments

Machine learning models stratified by environmental conditions can accurately predict depression risk among middle-aged and older adults while identifying context-specific patterns of feature attribution, according to a study analyzing nationwide longitudinal survey data. Researchers demonstrated that machine-learning algorithms trained separately across distinct ambient air-quality environments achieved good discriminatory power in identifying individuals at risk of depression. Although the relative importance assigned to specific predictors shifted across cleaner and more polluted environments, formal statistical testing revealed that these differences did not constitute a statistically significant interaction between individual risk factors and ambient air quality.

By examining participant records documented within this longitudinal cohort, the research team evaluated how broader environmental circumstances influence algorithmic assessments of psychological well-being. The analytical design focused specifically on predicting depression risk across different air-quality contexts, investigating whether machine-learning architectures could capture differing risk profiles when applied to populations living under varying ambient environmental conditions.

To establish distinct environmental contexts for model training, the authors evaluated ambient air quality using the annual average Air Quality Index. The study stratified participant records according to the policy-relevant threshold of 100 defined by the Chinese Air Quality Index framework. This regulatory benchmark divided the study population into two comparative environmental categories: individuals residing in areas with an annual average index of 100 or below, representing acceptable air quality, and those exposed to an index exceeding 100, corresponding to unhealthy air quality. This stratification enabled the team to develop independent predictive models for each environmental category.

The research team developed and trained four distinct machine learning models to predict depression risk within each air-quality stratum: Extreme Gradient Boosting, Gradient Boosting Decision Trees, Gaussian Naive Bayes, and Random Forest classifiers. Each model was specifically calibrated to predict depression risk separately in the acceptable air-quality group and the unhealthy air-quality group.

Model performance across the evaluation pipeline was assessed using the area under the receiver operating characteristic curve to quantify discriminatory capability in each air-quality group. The models demonstrated good discriminatory power overall across both environmental strata, confirming the feasibility of machine-learning approaches for depression-risk prediction. Across the evaluated architectures, ensemble tree-based methods achieved the strongest results, with Extreme Gradient Boosting and Gradient Boosting Decision Trees attaining the highest area under the curve values in distinguishing between presence and absence of depression risk.

To examine the internal decision-making mechanics of the resulting models, the investigators applied SHapley Additive exPlanations analysis, characterizing feature-attribution patterns within each separate environmental stratum. The interpretability analysis identified both shared baseline predictors and distinct shifts in relative feature weighting between the two groups. Life satisfaction and memory ability were prominent predictors of depression risk across both environmental strata. In contrast, self-rated health and reports of bodily pain received relatively greater feature attribution within the predictive model trained on the unhealthy air-quality cohort.

Following the machine-learning attribution phase, the researchers conducted post-hoc interaction analyses to formally examine whether the associations between individual predictors and depression risk differed significantly across the air-quality strata. In formal statistical evaluation, the omnibus predictor-by-air-quality interaction failed to reach statistical significance, yielding a P-value of 0.220. Furthermore, none of the predictor-specific interactions remained statistically significant after the application of multiple-testing correction, indicating that the observed differences in feature attribution did not reflect robust statistical effect modification between predictors and air quality.

Despite the lack of formal statistical effect modification, the authors concluded that air-quality-stratified machine-learning models showed good predictive performance and revealed context-specific differences in model-derived feature attribution. While conventional interaction analyses did not identify statistically robust modification, the stratified SHapley Additive exPlanations findings provide complementary information on how depression-risk models may rely on different combinations of individual characteristics across environmental contexts. The findings demonstrate that interpretable machine-learning frameworks can successfully identify depression risk while illuminating subtle differences in how individual indicators function across disparate environments.

Subject of Research: Machine learning prediction of depression risk stratified by air quality

Article Title: Air quality-stratified interpretable machine learning for depression risk in middle-aged and older adults

Article References: Li, R., Jia, L., Qiu, Y., & Song, X. (2026). Air quality-stratified interpretable machine learning for depression risk in middle-aged and older adults. Air Quality, Atmosphere & Health, 19(11), Article 265. https://doi.org/10.1007/s11869-026-02157-4

Image Credits: AI Generated

DOI: 10.1007/s11869-026-02157-4

Keywords: China Health and Retirement Longitudinal Study, CHARLS, air quality index, Chinese AQI framework, AQI stratification, depression risk, middle-aged and older adults, interpretable machine learning, XGBoost, gradient boosting decision tree, random forest, Gaussian Naive Bayes, area under the curve, SHAP analysis, feature attribution, life satisfaction, memory ability, self-rated health, body pain, interaction analysis