Kaiser Permanente Suicide-Risk Model

Kaiser Permanente Northern California's suicide-risk model is a machine-learning model built into its electronic health record. It scores patients at intake to a large virtual mental-health program. When a score passes a pre-set threshold, a dashboard flag reminds the intake clinician to assess suicide risk.

How it is used

The program handles more than 5,000 intake visits a month. An internal predictive-analytics core computes the score in near real time. Only set kinds of visits trigger a score, and it is ready about 30 minutes later.

The flag joins an alert that already existed: a positive self-report screen. The screen means two questionnaires, the PHQ-9 (Patient Health Questionnaire-9) and the C-SSRS (Columbia-Suicide Severity Rating Scale).

Either alert starts the same workflow. The clinician is reminded to conduct a suicide-risk assessment at the upcoming visit. If the patient does not attend, up to two more attempts are made to contact them (outreach).

The team chose intake as the point of care because, by its own account, over 70 percent of treatment disengagement happens after the first or second visit. The sources read for this case do not corroborate that figure independently.

Who decides

The flag decides nothing. It is advisory and never gates care. The clinician conducts the suicide-risk assessment and decides any response.

The sources publish no rate at which clinicians accept or override the flag, or complete the assessment.

What the added flag changes

Either alert alone starts the workflow. So the model can add patients to the workflow but can never remove one. The two alerts are added together, not compared.

The event the model predicts is rare. In the validation study, 0.17 percent of intake appointments were followed by a suicide attempt within 90 days. Among the tenth of appointments the model ranked riskiest, 0.8 percent were.

So the large majority of flagged patients will not attempt in that window. At more than 5,000 intakes a month, the flag adds many false alarms and much clinician work. The implementation team itself raised that caution.

What the validation found

Papini and colleagues published the validation in JAMA Psychiatry in 2024. They studied 1,623,232 intake appointments from 835,616 patients, from January 2012 to April 2022. Of those appointments, 2,800, or 0.17 percent, were followed by a suicide attempt within 1 to 90 days. Of those attempts, 78, or 3.9 percent, were fatal.

The model's area under the curve was 0.77. That measure shows how well a score ranks a patient who later attempts above one who does not, where 0.5 is chance and 1 is perfect.

Sensitivity is the share of real cases the model flags. Specificity is the share of other patients it leaves unflagged. Sensitivity was 37.2 percent at 95 percent specificity, and 18.8 percent at 99 percent specificity.

The tenth of appointments ranked riskiest held 48.8 percent of the appointments later followed by an attempt. Its positive predictive value, the share of those appointments followed by an attempt, was 0.8 percent.

Ranking accuracy varied by group. The area under the curve ranged from 0.69 for one race group to 0.89 for another. It was 0.79 for women and 0.74 for men. The team cited this as an equity consideration.

TechTarget's HealthTech Analytics restated these figures in 2024. That was a restatement of the study, not a separate test.

How it was built

Kaiser Permanente's Division of Research and data-science staff at The Permanente Medical Group built and validated the model. It draws on earlier work by the Mental Health Research Network.

The implementation followed a five-step community-engaged process. First, the team assessed whether the model worked as well for each demographic group. Second, it interviewed patients with lived experience of suicide-related thoughts. Third, it gathered clinicians' views. Fourth, it assembled the workflow with operational leaders. Fifth, it educated and tested, in repeated rounds.

An institutional review board, which is an ethics review board, approved the work. It ran under a Delivery Science Grant. The authors stress that deployment must respect patient preferences, avoid reinforcing inequities, and manage workforce demands.

What the evidence does and does not show

The published record is a feasibility and design report, not an effectiveness study. It documents the workflow and its rationale. It presents no evaluation showing that the deployment reduced suicide attempts.

The peer-reviewed case study appeared in NEJM Catalyst Innovations in Care Delivery in February 2026. A medRxiv preprint from April 2025, which is not peer reviewed, mirrors it.

A related Mental Health Research Network study across four health systems found that models of this kind kept their ranking accuracy despite changes in how care was delivered.

Which record system

Kaiser Permanente Northern California runs KP HealthConnect, an Epic system, and the sources describe dashboards built within the record. The peer-reviewed text says electronic health record and names no vendor. So naming Epic is well-supported context, not a claim the study makes.

No large language model or generative AI is involved. This is a regression model over structured record fields. Reading it as a generative AI tool overstates it.

Where to look

Watch the two alerts that meet at the intake clinician: Model flag to the clinician's dashboard and Screen alert to the clinician. The case file's defining gap is that no standing check compares them.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. Closing a pathway does not stop the work along it. It means mistakes stop passing along it.

Self-correcting means the network's mistakes are corrected rather than building on each other. At a tipping point, they are close to building on each other.

This case has a budget of 9 units. Explore (No Targets) sets no targets.

Under Service Targets Only, one tool is enough to meet the targets. The cheapest each cost 2 of the 9 units: Keep skills sharp, Escalate checks, Review on schedule, or Mark AI-written records.

Under Service and Safety Targets and All Governance Targets, the targets can also be met. Both levels ask you to close every failure pathway, among other targets. The cheapest way costs 7 of the 9 units and uses three tools: Escalate checks, Mark AI-written records, and Store less data.

At both levels, eight distinct combinations fit the budget, counting stronger settings. Every one of them includes those three tools.

Escalate checks stops mistakes passing along Model flag to the clinician's dashboard. Mark AI-written records stops them passing along four pathways. They are History read at the assessment, Record data read by the model, Screen alert to the clinician, and Screen answers as model input. Store less data stops them passing along Assessments written to the health record.

More is not better here. Every tool at its standard setting costs 21 units, and at its strongest setting 36. Either way every failure pathway closes, but the model is no longer clearly helping the work. So using everything misses the targets at every level that sets them.