Harvard Researchers Predict 75% of Suicide Attempts One Week Before They Occur
Researchers at Harvard's Nock Lab have predicted 75% of suicide attempts and 87% of suicide-related events in the week before they happened, using a multiyear study of more than 600 high-risk adults and adolescents. The findings appear in the October issue of the Journal of Psychopathology and Clinical Science (DOI 10.1037/abn0001117), as reported by Harvard's Faculty of Arts and Sciences publication Current on September 8, 2026.
The study was led by Matthew K. Nock, the Edgar Pierce Professor of Psychology at Harvard University and a licensed clinician. Nock, a 2011 MacArthur Foundation fellow, has spent his career building the empirical foundations for what the field now calls real-time suicide risk forecasting. He pioneered the use of ecological momentary assessment (EMA) — a method of repeatedly sampling a person's thoughts and feelings in real time, originally done via personal digital assistants — to establish that suicidal ideation fluctuates moment to moment rather than building along a steady upward path.
Study participants came from two clinical populations: adults who had received emergency department-based psychiatric treatment, and adolescents aged 12 to 19 treated at an inpatient clinic for suicidal thoughts or behavior. The recruitment design targets the highest-risk strata, where prediction has the most clinical value but also where false positives carry the greatest potential for harm.
Suicide is the second-leading cause of death for Americans aged 10 to 34, trailing only accidents. Approximately 50% of people who died by suicide saw a clinician in their final month, according to multiple studies cited in the Harvard report. That statistic points to the clinical gap the Nock Lab's work tries to address: clinicians are already in contact with at-risk patients, but lack reliable near-term forecasting tools to act within the window that matters most.
The timing is notable. A separate Nock Lab paper, "Describing and Measuring the Pathway to Suicide Attempts," found that 86.5% of proximal planning steps took place within one week of attempting and 66.6% occurred within 12 hours. The one-week prediction window in the current study aligns directly with the period in which most actionable planning crystallizes.
The Nock Lab has been building toward this result across multiple complementary tracks. A 2026 peer-reviewed review by Nock and Wang in Current Directions, titled "Understanding, Predicting, and Preventing Suicide," surveyed advances in technology-based methods for prediction and prevention. A preprint by G. Hang and colleagues (arXiv:2511.18199v1, released November 2025), "Improving Forecasts of Suicide Attempts for Patients with Little Data," tackled the cold-start problem: predicting attempts for patients with sparse longitudinal EMA data, which is a structural challenge given the rarity of suicide attempts even in high-risk cohorts. A separate Nock Lab effort developed a risk prediction model using readily available electronic health record data to forecast suicide attempts or death by suicide, broadening the approach beyond intensive EMA collection.
The convergence matters. EMA-based models capture high-frequency, within-person signal; electronic health record (EHR)-based models leverage data already sitting in clinical systems. Neither alone covers the full prediction landscape, and the Hang et al. preprint directly confronts the data sparsity that limits individual-level forecasting. Together, these threads suggest a field moving toward layered prediction architectures rather than a single model class.
The broader context here is that the clinical translation path for predictive suicide models is not straightforward. The base rate of suicide attempts, even in high-risk clinical samples, is low enough that specificity becomes the binding constraint. A model that identifies 75% of attempts is clinically meaningful only if its false-positive rate does not overwhelm the intervention resources it triggers, or cause harm through unnecessary hospitalization or crisis response. The study's recruitment from emergency and inpatient psychiatric settings means the reported performance figures apply to an acutely elevated-risk population, and generalization to outpatient or community samples is not established by these results.
The regulatory and ethical infrastructure for deploying algorithmic suicide risk prediction in routine care also remains immature. Clinical decision support tools that trigger high-stakes interventions based on probabilistic forecasts require validation frameworks, liability structures, and patient consent models that the field has not yet fully developed. The Merck Propecia case, in which newly unsealed court documents revealed that the pharmaceutical company and U.S. regulators had reports of suicidal behavior in men taking the anti-baldness drug, illustrates how slowly regulatory systems can respond to psychiatric risk signals even when the data exists.
Broader risk factors remain consistent across populations. A study reported by Reuters found that, regardless of geography, people who are young, single, female, poorly educated, or mentally ill are at higher risk. National data reported by Reuters in 2013 indicated that about one in 25 U.S. teens had attempted suicide and one in eight had thought about it.
What makes the Nock Lab result difficult to dismiss is the combination of temporal specificity, sample size, and the convergence with the lab's own pathway data. The finding that most proximal planning occurs within a week of an attempt, paired with a model that predicts 75% of attempts within that same window, suggests the field may be approaching a clinically actionable forecasting horizon. Whether health systems can build the deployment infrastructure to use it is a separate and harder question.
