One item on a depression questionnaire decides who gets assessed today
- The study and its design. A quality improvement evaluation of the Zero Suicide model across six US health systems in California, Oregon, Washington, Colorado and Michigan, using an interrupted time series over January 2012 to December 2019, published in JAMA Network Open in April 2025 by Ahmedani and colleagues. The population was patients aged 13 and over with an outpatient visit to a mental health specialty practitioner who had been members of the affiliated health plan for at least 10 of the 12 preceding months. Median 309,107 unique patients per month, range 55,354 to 451,837. There was no randomisation and no control arm; four systems adopted the model inside the observation window and two had adopted it before it, in 2012 and 2001.
- The route, in the order it runs. All six systems used the ninth item of the Patient Health Questionnaire-9 as the primary suicide risk screening question at outpatient mental health visits. A positive screen leads to a structured suicide risk assessment, the Columbia Suicide Severity Rating Scale at five systems and a locally built assessment of evidence-based factors at Henry Ford Health. A positive assessment leads to brief intervention, and from there to outpatient care including evidence-based psychotherapy.
- How much of the route was walked in 2019. Of 4,674,515 eligible mental health visits, 3,416,904, or 73 percent, included a suicide risk screening. Of 128,996 individuals who screened positive, 99,098, or 77 percent, received a risk assessment. Of 19,528 individuals with a positive risk assessment, 16,094, or 82 percent, received a safety plan, means counselling, or both.
- The primary outcome and what happened to it. The outcome was a composite standardised monthly rate of suicide attempt per 100,000 patients, counting any encounter coded as intentional self-harm together with death coded X60 to X84 or Y87.0, occurring within 90 days of the index month. Baseline rates were at least 30 to 40 per 100,000 patients per month at every implementing site and fell below 30 at three sites by 2019. Slope estimates after implementation were 0.69 lower per 100,000 per month at system A (SE 0.33, P = .04), 0.65 lower at system B (SE 0.33, P = .05) and 0.13 lower at system C (SE 0.04, P = .003), while system D moved 0.23 in the opposite direction (SE 0.24, P = .35). Every step-change term at the moment of implementation was non-significant, with P values of .43, .55, .18 and .29.
- The secondary outcome. For quarterly suicide death rates, the slope after implementation was 0.20 lower at system B (SE 0.06, P = .001) and 0.08 lower at system C (SE 0.03, P = .009); systems A and D showed no change (P = .19 and P = .46). The system that adopted the model in 2001 held the lowest sustained rate throughout, moving from 11.3 to 0.3 per 100,000 patients per month.
The first minutes of the appointment carry one administrative act. Before the conversation has taken any shape, the clinician has the patient's PHQ-9 in front of them and reads the ninth item, the line about thoughts of self-harm. In all six health systems evaluated here, spread across California, Oregon, Washington, Colorado and Michigan, that item was the primary suicide risk screening question, and it was put at the outpatient mental health visit itself.
The item is a self-report line lifted from a depression questionnaire. It sits above a base rate of at least 30 to 40 suicide attempts per 100,000 patients per month, counted within a 90-day window after the visit. On that arithmetic no single question separates the people who will attempt from the people who will not, and the Zero Suicide model does not ask it to. What a positive answer buys is not a probability. It buys an obligation.
That obligation is the whole of the design. A positive screen is followed by a structured risk assessment, the Columbia Suicide Severity Rating Scale at five of the systems and a locally built assessment at Henry Ford Health. A positive assessment sends the patient on to brief intervention, and from there to outpatient care that includes evidence-based psychotherapy. What was implemented was the requirement that each step follow the last. The accuracy of the opening question was left exactly where it was.
How far patients actually travelled down that route is counted separately. Of 4,674,515 eligible mental health visits in 2019, 3,416,904, or 73 percent, included a screening. Of the 128,996 people who screened positive, 99,098, or 77 percent, received a risk assessment. Of the 19,528 whose assessment came back positive, 16,094, or 82 percent, received a safety plan, means counselling, or both. Compounded across the three steps, roughly one in eight of those who answered yes reached the last of them.
The rates moved slowly where they moved. Three implementing sites were below 30 per 100,000 patients per month by 2019. Implementation was associated with a change in slope at three of the four systems that adopted the model inside the observation window, by 0.69 (SE 0.33, P = .04), 0.65 (SE 0.33, P = .05) and 0.13 (SE 0.04, P = .003) per 100,000 patients per month, while the fourth showed no change (P = .35). Every step-change term at the moment of implementation was non-significant. Nothing switched off; a line bent.
The authors are explicit that the model went in at system level and that the contribution of any single element of the route could not be estimated from these data. What the study describes is a rule about sequence, and that rule is what a clinician inherits at the desk: ask at the visit, and treat a positive answer as the reason the next step follows.
Of 4,674,515 eligible visits in 2019, 73 percent carried a screening; of the 128,996 positive screens, 16,094 people reached the brief intervention step.
This was a quality improvement evaluation with an interrupted time series design, not a randomised trial, and the authors name unmeasured bias as a consequence: each system decided for itself when and how to adopt the model. Pre-implementation trends were not alike, which matters for reading the slopes; system B was already falling before adoption (0.11 per 100,000 per month, P < .001) and system A was rising (0.44, P < .001). A difference-in-differences analysis was considered and could not be used, because the parallel trends assumption was violated. Results are at system level only, and individual-level effects could not be assessed, so the funnel figures for 2019 are a fidelity snapshot and are not linked to outcomes patient by patient. The contribution of any single step in the route cannot be separated out. Six large health systems with insured populations may not represent other settings, and 69.9 percent of the July 2017 population held commercial insurance. The observation period spans the ICD-9 to ICD-10 transition, which the authors expect to have undercounted attempts in the earlier years and therefore to have made the estimates conservative. The outcome window was fixed at 90 days and another window may give another answer. Power was limited for suicide death at systems A and D. The source reports no timings for any step of the route, so nothing here says how long a screen, an assessment or a brief intervention takes.