Abstract
BACKGROUND: Passive smartphone sensing shows promise for suicide prevention, but behavioral metadata (GPS, screen time, and accelerometry) often lacks the contextual information needed to detect acute psychological distress. Analyzing what people actually see, read, and type on their phones-rather than just usage patterns-may provide more proximal signals of risk. OBJECTIVE: This study aimed to test whether vision-language models (VLMs) applied to passively captured smartphone screenshots can predict momentary suicidal ideation (SI). METHODS: Seventy-nine adults with past month suicidal thoughts or behaviors completed ecological momentary assessments (EMA) over 28 days while screenshots were captured every 5 seconds during active phone use. We fine-tuned open-source VLMs (Qwen2.5-VL [Alibaba Cloud], LFM2-VL [Liquid AI]), and text-only models (Qwen3 [Alibaba Cloud]) to predict SI from screenshots captured in the 2 hours preceding each EMA. We evaluated performance with temporal and subject holdouts. RESULTS: The analytic sample comprised 2.5 million screenshots from 70 participants. Temporal holdout models achieved strong discrimination at the EMA level (AUC=0.83; AUPRC=0.77), with image-based models outperforming text-only models (AUC=0.83 vs 0.79; 95% CI 0.003-0.07). Subject holdout generalization was near chance (AUC≈0.50), though a simple lexical screening method retained modest discrimination (AUC=0.62). Smaller models performed comparably to larger models, supporting feasible on-device deployment. CONCLUSIONS: Screen content predicts short-term SI with clinically meaningful accuracy when models are personalized but does not generalize across individuals. These findings support a 2-stage clinical architecture, coarse lexical screening for new patients, with personalized VLM-based monitoring after a calibration period. On-device inference may enable privacy-preserving deployment.