Conference Agenda

This introductory talk reviews the theoretical foundations and empirical evidence supporting precision stratification in suicide risk assessment. Professor Courtet synthesizes findings from his landmark studies on biological markers (inflammation, neuroimaging), psychological dimensions (anhedonia, psychic pain), and social factors (interpersonal theory constructs). Key results from large French cohorts demonstrate how multimodal risk profiles outperform traditional approaches, identifying latent subgroups with distinct trajectories. The presentation establishes the conceptual framework for passive monitoring applications presented in subsequent talks, highlighting precision psychiatry's potential to transform suicide prevention from reactive to proactive paradigms.

Adolescence is a developmental period marked by heightened vulnerability to suicidal thoughts and behaviors. Digital monitoring offers new opportunities for continuous, ecologically valid assessment of daily experiences. Prior work has shown that digital self-reports, particularly high-frequency ecological momentary assessment (EMA), enhance detection of short-term changes in suicidal ideation, affective instability, and interpersonal stressors. In parallel, passive smartphone sensing, including usage patterns, mobility, sleep-related indicators, and social media engagement, provides an unobtrusive way to capture behavioral signatures that may precede shifts in risk. This talk presents data from a cohort of 99 adolescents receiving care in a specialized depression and self-harm clinic who participated in a six-month digital monitoring protocol that combined continuous passive smartphone tracking (via the iFeel app) with weekly self-reports of suicidal ideation and behavior. We examine how adolescents' total screen time and average daily social media use across each preceding week relates to self-reported suicidal ideation and risk, and how baseline depression and peer functioning shape these associations. Mixed-effects models revealed no main effect of total screen time and social media exposure on suicidal outcomes. However, between-person depressive symptoms and peer functioning moderated the direction of associations between social media usage and suicidal ideation, indicating that passive sensing signals must be interpreted in light of individual vulnerabilities and interpersonal context. Aligned with emerging evidence in digital phenotyping, these findings show that integrating active self-report and passive behavioral sensing data can improve temporal precision in suicide-risk monitoring for adolescents. This approach may support scalable, continuous, and context-sensitive suicide-prevention strategies in adolescents.

Military personnel experience alarming rates of suicide deaths that have been slowly increasing over time. A variety of factors contribute to increased risk in this population, but we have limited ability to predict when service members are transitioning from a relatively lower to a higher risk state. Most prior prediction research has focused on suicidal ideation, which does not translate into strong prediction of suicidal behavior. Continuous wearable device monitoring might improve the ability to predict suicide-related events among military personnel, though this has not yet been explored in the literature.

Active-duty service members and Veterans (N = 90) were provided with a Fitbit Inspire 2 and instructed to wear it continuously (except charging) for 28 days. Heart rate, heart rate variability (high frequency, low frequency, and RMSSD), physical activity (including calories burned, steps, and intensity), and sleep parameters were continuously captured and were entered into machine learning (heterogeneous mixture and change-point detection, HetMM-CPD) pipeline to predict daily ecological momentary assessments for suicide attempts and non-suicidal self-injurious behavior.

The machine learning pipeline achieved robust discrimination for suicide and self-injurious-related events (AUC = .90). The best predictive sources were physical activity (steps, calories) and heart rate variability (RMSSD). Sleep variables did not improve performance. At least 3 days of continuous data were required to predict subsequent suicide events.

Continuous observation of U.S. military service members with a consumer-grade, affordable wearable device can result in sensitive and specific prediction of suicide and self-injurious-related events. These findings suggest the possibility that in the future, continuous monitoring could flag the need for interventions to be deployed in real-time to interrupt the transition from lower- to higher-risk states among service members. This machine learning pipeline offers the opportunity to transition away from predicting only suicidal thinking into predicting, and potentially interrupting, suicidal and self-injurious behavior.

Short-term prediction of suicidal crises remains imprecise when relying solely on clinical interviews and traditional risk factors, limiting timely preventive action. This study evaluated whether real-time, sensor-based monitoring of daily behavior patterns via patients’ own smartphones could predict, within 7 days, suicide attempts or emergency psychiatric visits in a high-risk cohort with suicidal ideation.

A total of 225 outpatients with a history of suicidal thoughts or behavior were recruited within the multicenter SmartCrisis study and followed for 6 months. The eB2 app passively collected data on distance walked, steps, time at home, and app use, building individualized daily activity profiles and using an unsupervised Bayesian change-point algorithm to detect abrupt shifts in behavioral pattern distributions over time.​

During follow-up, 18 participants (8%) attempted suicide and 14 (6.2%) required emergency department psychiatric care, yielding 32 suicidal risk events. Behavior changes detected by the algorithm predicted suicide risk within a 1‑week window with an area under the curve of 0.78, indicating good discriminative performance in real-world clinical conditions.​

Combining passive sensing, individualized digital phenotyping, and online change detection offers a scalable way to flag imminent suicidal crises and could support stepped-care intervention models that intensify resources when a patient’s behavior deviates from their usual pattern. Integrating these digital signals with ecological momentary assessment, electronic health records, and clinical evaluation may enhance short-term suicide risk detection and is potentially transferable to other psychiatric conditions where early crisis prediction is crucial.