Harvard study predicts 75% of suicide attempts a week in advance

A Harvard-led team has predicted 75 percent of suicide attempts during the week before they happened, using a model built from smartphone mood surveys completed several times a day by people recently discharged from psychiatric care. The multiyear study, which followed more than 600 high-risk adults and adolescents, also anticipated 87 percent of what the researchers call suicide-related events, a category covering both attempts and hospitalizations carried out to prevent them. The findings are forthcoming in the October issue of the Journal of Psychopathology and Clinical Science, according to a report from Harvard's Faculty of Arts and Sciences.

From year-long horizons to a single week

Suicide prediction has long relied on models that estimate a person's risk over the coming six months, year or even decade, drawing mainly on retrospective self-reports asking whether someone attempted suicide since the previous survey. That framing clashes with how risk actually behaves. Matthew K. Nock, the Edgar Pierce Professor of Psychology who led the work, has spent his career showing that suicidal thoughts fluctuate from moment to moment, while clinical care is organized around scheduled appointments. The gap is stark: across multiple studies over recent decades, including Nock's own, half of the people who died by suicide had seen a clinician during their final month. Suicide is also the second-leading cause of death for Americans aged 10 to 34, behind only accidents.

How the data was collected

Two groups participated: adults who had received emergency-room psychiatric treatment, and 12-to-19-year-olds treated at an inpatient clinic for suicidal thoughts or behavior. Starting immediately after release, both groups received optional app-based surveys — six per day for the first three months, then one per day for the following three. Each survey posed 20 questions answered on 0-to-10 sliders: three measuring aspects of suicidal thinking (urge, intent, and perceived ability to resist urges) and 17 covering emotional states such as hopelessness, isolation, anger, agitation, worry, fatigue and positivity. Close to 500 unique participants completed at least one full survey, producing more than 77,000 responses in total. Beyond the answers themselves, the team also analyzed behavioral metadata, including how quickly a person opened a survey after being prompted and how long they took to finish it.

Nock, who received a MacArthur Foundation fellowship in 2011, pioneered this kind of real-time surveying in the era of personal digital assistants before smartphones. The new study shows the approach can feed an actual short-horizon prediction system, not just descriptive research.

Agitation outweighed depression as a warning sign

Among the emotional states measured, agitation was a far stronger indicator of risk than depression: every additional point of self-reported agitation on the slider corresponded to an 11 percent increase in the likelihood of a suicide attempt. That lines up with other recent work from Nock's lab, in which 90 percent of people who survived an attempt said they had been seeking relief from psychological agitation and pain that, in hindsight, felt temporary.

A system built for intervention, not just prediction

The study ran with a real-time alert system that triggered a clinical response when a participant reported high suicidal intent. Nock framed the goal as building something scalable, reproducible and accurate enough to deliver support before a crisis rather than after it. The same research program extends to wearable sensors that track sleep, heart-rate variability, skin conductance and voice signatures, and to just-in-time interventions that prompt a person to use coping techniques learned in therapy or to reach out to a clinician or loved one at moments of elevated risk. Nock compared the approach to monitoring blood sugar or cardiovascular risk, while acknowledging that mental-health monitoring involves extra privacy considerations and must be developed together with clinicians and the people being treated. The research was partially supported by federal funding from the National Institute of Mental Health.

Why it matters

The statistic that half of suicide deaths had recent clinical contact has usually been read as evidence that warning signs are invisible. This study suggests the opposite: the signals exist, but they move too fast for appointment-based care to catch. If risk can be detected days ahead rather than estimated over years, resources can shift toward timely outreach at the moments it matters. Real-world deployment still has open questions — the participants were a high-risk, post-discharge population, and sustained use would require managing survey fatigue, alert handling and sensitive data at scale. Even so, the results demonstrate that short-horizon prediction of suicide attempts is achievable, which changes what clinicians can reasonably expect from risk assessment tools.