Classical machine learning versus transformers for predicting daily psychological distress from passive smartphone sensing and self-report: a prospective cohort study

In a prospective cohort of 169 community-dwelling adults, a Platt-calibrated SVM predicted daily significant psychological distress (BSI-18 GSI T-score ≥63) with AUROC 0.84 (95% CI 0.71–0.93).

• Why it matters: Daily passive smartphone sensing plus end-of-day self-report could enable clinically useful triage, but it is unclear whether classical models match transformers or whether multimodal architectures add value.
• What they did: Participants provided 7 days of passive sensing (accelerometer, GPS, power state) and BSI-18 ratings; days with distress totaled 1,062 with 11.2% prevalence, and eight tuned models were evaluated on held-out participants (210 records, 25 cases) with participant-level bootstrap 95% CIs.
• The result: Best discrimination came from the SVM (0.84), the multimodal fusion transformer (0.84), and XGB (0.82), while sensor-only transformers were at chance (0.49–0.50); elastic net logistic regression had the top sensitivity (0.56) and PPV (0.64) at threshold 0.50.

AI summary of the abstract · may contain errors · not medical advice · read the paper.