AI suicide risk prediction models can rank patients by risk better than chance, but they cannot reliably predict who will die by suicide. A 2025 PLOS Medicine meta-analysis of 53 studies concluded they are too inaccurate for screening or for allocating treatment. Their value depends entirely on what care follows a flag.
Suicide prediction is one of the most emotionally compelling uses of machine learning in medicine. The logic seems unanswerable: many people who die by suicide had contact with health services in the months before. The data were in the record. If an algorithm could read those records and raise a hand, surely lives would be saved.
That logic has driven more than a decade of research, several national programs and a great deal of vendor marketing. The results are now clear enough to state a thesis that sounds harsh but is actually hopeful. Prediction is not prevention. Because suicide is rare, even a good model will produce mostly false alarms, and a flag without a response saves no one. The systems that show benefit are the ones that treat the algorithm as a way to deliver more care to more people, not as a way to find the one person who is about to die.
If you are thinking about suicide or worried about someone, please contact your local emergency services or a crisis line now. In the US, call or text 988. Help is available, and talking to someone can make a difference.
How accurate is AI at predicting suicide?
The most comprehensive answer arrived in September 2025. Matthew Spittal and colleagues at the University of Melbourne, with co-authors including Keith Hawton and Nav Kapur, published a systematic review and meta-analysis in PLOS Medicine covering 53 studies of machine learning algorithms that predicted suicide or hospital-treated self-harm.
The models’ areas under the curve ranged from 0.69 to 0.93, which sounds impressive. But the numbers that matter at the bedside were less flattering. Sensitivity ranged from 45% to 82%, meaning that in many studies more than half of the people who went on to self-harm or die by suicide had been classed as low risk. Specificity was high, 91% to 95%, but in a rare outcome that still produces a flood of false positives. Using the prevalence in the original study samples, positive predictive values ranged from 6% to 17%. In low-prevalence populations, the authors estimated it could fall to 0.1%.
Their conclusion was blunt: the accuracy of these algorithms is too low to be useful for screening or for prioritizing high-risk individuals for interventions.
This was not a new finding so much as a confirmation. In 2019, Ross Belsher and colleagues reviewed suicide prediction models in JAMA Psychiatry, including those being developed by the US Department of Defense, the Department of Veterans Affairs and Kaiser Permanente, and found positive predictive values remained low even at the highest risk thresholds. The statistical problem is structural. It is not a matter of waiting for better models.
Clinical guidance has already absorbed this lesson. NICE’s 2022 guideline on self-harm in England tells clinicians not to use risk assessment tools and scales to predict future suicide or repetition of self-harm, and not to use global low, medium or high risk labels to decide who gets treatment. Algorithms are a more sophisticated version of the same idea, and they face the same objection.
Why do suicide prediction models produce so many false alarms?
Because of base rates, and base rates are ruthless.
Take a hypothetical health system with 100,000 adult patients, in a year in which 100 of them, 0.1%, will make a serious attempt. Suppose a model catches half of them, a sensitivity of 50%, and correctly clears 95% of everyone else. It will flag 50 people who go on to attempt. It will also flag nearly 5,000 who do not. Of every 100 flags, roughly one belongs to a person who will make an attempt that year. Meanwhile, the other 50 people who attempt were never flagged at all.
Now imagine that system’s single suicide prevention nurse receiving 5,000 alerts.
This is why “accuracy” is the wrong question. A model can be statistically excellent and operationally useless if the response it triggers cannot scale to the number of flags, or if the response is so heavy, such as involuntary assessment, that applying it to thousands of false positives would cause harm of its own.
The same arithmetic flips the other way when the flagged group gets something low-cost, acceptable and helpful to almost anyone. A check-in call, a safety planning conversation or a faster appointment does little harm to the person who was never going to attempt, and may help many people who were struggling in ways the model did not measure. That is the design space where prediction can earn its keep.
Do suicide prediction programs reduce deaths?
The largest real-world test is the US Department of Veterans Affairs program REACH VET, which since 2017 has used an algorithm to identify veterans in the top 0.1% of predicted risk and prompts clinicians to review their care and reach out.
A 2021 evaluation by John McCarthy and colleagues in JAMA Network Open found that inclusion in the program was associated with more outpatient appointments, more new safety plans, fewer mental health admissions and emergency visits, and fewer documented suicide attempts. Those are meaningful gains in care.
It did not find a difference in suicide deaths. A follow-up study in JAMA Network Open in July 2025, covering implementation from 2017 to 2021, again found no significant association between program inclusion and suicide mortality at six, 12 or 24 months. The authors suggested additional prevention approaches were needed.
At Vanderbilt University Medical Center, Colin Walsh’s team took a different angle and tested the interface rather than the model. In a randomized trial published in JAMA Network Open in early 2025, their risk model’s flags were delivered either as interruptive pop-up alerts or as a passive score in the chart. Clinicians completed a suicide risk assessment after 42% of interruptive alerts but only 4% of passive ones. The algorithm was the same. The response was ten times different.
Taken together, these studies tell a consistent story. Models can change what clinicians do. Whether that changes outcomes depends on what the clinicians do next.
The strongest case for algorithmic risk flags
The counterargument is serious. The traditional alternative is also poor: a widely cited 2017 meta-analysis in Psychological Bulletin, led by Joseph Franklin, found that five decades of research on known risk factors predicted suicidal thoughts and behaviors only slightly better than chance. If both methods are weak, an algorithm at least applies the same standard to every patient, does not tire, and can scan an entire population that no clinician will ever see in one room.
REACH VET’s increases in outreach, safety planning and attendance are not trivial. For a veteran who would otherwise have fallen through the cracks, a phone call prompted by an algorithm may matter enormously, even if the effect is too small to show up in population mortality statistics over two years.
That case is correct that the alternative to algorithms is not perfect clinical intuition. Where it overreaches is in the framing. The value of these programs comes from extending care, not from forecasting. Programs designed around “finding the one” will disappoint. Programs designed around “reaching more people with something useful” can succeed even with imperfect models.
What this means for hospitals, clinicians and policymakers
For hospitals, the first question for any suicide risk tool should not be its AUROC. It should be: what happens to each flag, who does it, and can that response be delivered to every flagged person without crowding out other care? We made the same argument about capacity forecasting in Predicting Hospital Capacity Is the Easy Part. A forecast nobody can act on is a liability, and in this domain it can also be a legal and ethical one.
For clinicians, a flag should prompt a conversation, not a label. Evidence-based practices such as safety planning, follow-up contact after discharge and restricting access to lethal means through collaborative, respectful discussion help across the risk spectrum. A risk score should never be the reason someone is denied care because they scored “low”. Integrating this work into everyday care is the theme of Mental Health in Primary Care Works, where most people at risk are actually seen.
For policymakers and funders, the lesson is to fund the response, not only the model. An algorithm is cheap. The outreach nurses, crisis services and follow-up appointments it generates are not. As with diagnostic AI more broadly, discussed in AI Diagnostics: The Evidence Is Strongest Where the Question Is Narrowest, the useful question is narrower than the marketing.
For patients and families, the key message is that no algorithm replaces asking directly. If you are worried about someone, asking them about suicide does not put the idea in their head, and connecting them with help is what matters.
Frequently asked questions
Can AI predict suicide?
Not reliably at the individual level. A 2025 PLOS Medicine meta-analysis of 53 studies found machine learning models missed many people who later self-harmed or died by suicide and produced many false alarms. The authors concluded the models are too inaccurate for screening or allocating treatment. They may still help health systems target broadly useful outreach.
What is REACH VET?
REACH VET is a US Department of Veterans Affairs program, launched in 2017, that uses an algorithm to identify veterans in the top 0.1% of predicted suicide risk and prompts clinicians to review their care. Evaluations found more appointments, more safety plans and fewer documented suicide attempts, but studies published in 2021 and 2025 found no significant change in suicide deaths.
Why are suicide risk scores discouraged in clinical guidelines?
Because they predict poorly and can lead to people labelled “low risk” being denied help. NICE’s 2022 self-harm guideline advises against using risk tools or low, medium and high categories to predict future suicide or decide treatment. It recommends a needs-based assessment for everyone who self-harms instead.
What actually helps prevent suicide in health care?
Approaches with supportive evidence include safety planning, follow-up contact after emergency visits or discharge, treating underlying conditions such as depression and substance use, and collaborative conversations about reducing access to lethal means. These help across risk levels, which is why they pair better with imperfect prediction than intensive interventions reserved for a few.
Does AI suicide screening invade privacy?
It can raise real concerns, because models analyze sensitive records without patients necessarily knowing. Good practice includes transparency about whether such tools are used, clear governance over who sees flags, and ensuring a flag leads to supportive outreach rather than punitive action. Patients can ask their health system whether algorithmic risk tools are in use.
The takeaway
The value of suicide risk AI will never be in knowing who will die. It lies in giving health systems a reason, and a list, to reach more people with care that helps almost anyone. Build for the response, and the prediction becomes useful. Build for the prediction, and you have a very expensive worry.
If your organization uses a suicide risk algorithm, what happens in the hour after a flag? We would like to hear how it works in practice.
