**Background:** Psychological distress encompasses unpleasant emotions that impair daily functioning and interpersonal communication. Severe psychological distress causes moderate-to-serious impairment in social, occupational, or school functioning and requires treatment. While sociodemographic factors (age, gender, marital status, education) and health-related factors (chronic conditions, social media use, wearable device use) have been linked to mental health problems, prior work often relied on traditional machine learning with limited variable sets and subjective parameter selection. This study aimed to use auto-machine learning with broader feature selection to identify predictors of severe psychological distress from the HINTS database.
**Methods:** Data came from merged HINTS 5 Cycles 3 (Jan–May 2019, n=5438) and 4 (Feb–Jun 2020, n=3865). After removing missing data, the analytic sample was 4036 respondents. Severe psychological distress was measured using four items (two depression, two anxiety) from the PHQ-4; respondents answering "Nearly every day" or "More than half the days" to any item were classified as having severe distress. Fifty-seven predictor variables were initially selected covering sociodemographics, health status, and behavior. Chi-square and t-tests compared groups. Class imbalance was addressed with SMOTE. Five algorithms were applied: logistic regression (LR), Auto-GLM, Auto-Random Forests, Auto-Deep Learning, and Auto-GBM, all implemented in R 4.1.2 using the h2o package. Grid search with five-fold cross-validation optimized hyperparameters. Model performance was evaluated using accuracy, precision, recall, F1-score, and AUC.
**Key Results:** The sample included 739 (18.31%) individuals with severe psychological distress and 3297 (81.69%) without. Significant bivariate differences (p<0.05) included: female sex, lower income, not married, social media use, deafness, lack of friends/family to discuss health, and no moderate physical activity being more common in the distress group. Mean age was lower (47.30 vs. 53.40 years) and mean BMI (29.86 vs. 28.07), weight (188.74 vs. 179.76 lbs), and average sitting time (7.87 vs. 6.91 hours) were higher in the distress group. LR identified 40 significant predictors. Key ORs included: Hispanic/Latino origin (OR 2.28, 95% CI 2.00–2.61), deafness (OR 3.30, 95% CI 2.75–3.97), having a chronic disease (OR 2.68, 95% CI 2.40–3.00), believing nothing can lower cancer risk (OR 2.50, 95% CI 2.24–2.80), and ever using e-cigarettes (OR 2.07, 95% CI 1.81–2.37). Protective factors included higher education (OR 0.37, 95% CI 0.33–0.42), higher income (OR 0.23, 95% CI 0.20–0.26), and confidence in taking care of health (OR 0.07, 95% CI 0.05–0.09). Across all five algorithms, six variables were selected by every method: HaveDevice_Cat, GeneralHealth, OwnAbilityTakeCareHealth, IncomeRanges, IncomeFeelings, and Age. Auto-GBM outperformed all other models: accuracy 89.75%, precision 89.68%, recall 89.31%, F1-score 89.48%, and AUC 95.57%. LR had the lowest performance (AUC 79.99%).
**Clinical Implications:** The study demonstrates that auto-machine learning, particularly Auto-GBM, can effectively predict severe psychological distress using readily available survey data. The 47 identified predictors span sociodemographic, health, and behavioral domains, with novel findings around cancer information engagement and health confidence. These results can inform screening tools for healthcare providers and psychological counseling services, enabling earlier identification of at-risk individuals. The automated parameter tuning reduces subjectivity compared to traditional ML approaches. Limitations include potential multicollinearity among variables, lack of mechanistic analysis, and reduced interpretability of ensemble methods. Future work should explore deeper causal mechanisms and combine these predictors with explanatory methods.