A web-based decision tool helped more adults with major depressive disorder remain on their prescribed antidepressant through eight weeks in an international randomized trial. The result supports a role for PETRUSHKA in shared prescribing decisions, but it does not show that an algorithm can identify a guaranteed best drug for every patient.
The JAMA trial enrolled 520 eligible adults at 47 sites in Brazil, Canada and the United Kingdom. Participants were 18 to 74 years old, had major depressive disorder and were willing to use one antidepressant. Recruitment and follow-up ran from November 2022 to January 2025, and the findings were published online on March 4, 2026.
PETRUSHKA combines evidence from trials and health records with clinical and demographic information. Patients also enter preferences about adverse effects. The system then ranks three antidepressant options for discussion with a clinician. It is a decision-support system, not an autonomous prescriber, and the clinician and patient make the final choice together.
Early Treatment Continuation Was the Main Result
The primary outcome was whether participants stopped or changed the assigned antidepressant for any reason by week eight. Among the 493 people included in the main efficacy analysis, 41 of 241 in the PETRUSHKA group stopped treatment, compared with 69 of 252 in usual care. Those proportions were 17% and 27%, producing an adjusted relative risk of 0.62.
Stopping because of adverse effects was also less frequent at eight weeks: 9% in the PETRUSHKA group and 16% under usual care. These figures matter because all-cause discontinuation can reflect tolerability, perceived benefit and a patient's willingness to continue. The measure does not isolate which of those factors caused the difference.
The tool may have helped through more than one route. Its ranked options may have produced a better fit for some patients, while the structured conversation about side effects and preferences may itself have increased confidence in the selected treatment. The trial was not designed to separate the prediction component from the shared-decision process.
Twenty-Four-Week Findings Need Careful Reading
At 24 weeks, participants using PETRUSHKA reported lower depression and anxiety scores than those receiving usual care. The adjusted difference on the nine-item Patient Health Questionnaire was 1.92 points, while the difference on the seven-item anxiety scale was 1.39 points. Both comparisons favored PETRUSHKA.
The longer-term picture was less uniform. Observer-rated depression and anxiety scales did not show a meaningful difference at 24 weeks. Treatment discontinuation by that point was 34.4% with PETRUSHKA and 40.4% with usual care, a difference that was not statistically significant. The study also found no clear group difference in medication adherence, functional impairment or suicidality.
Two serious adverse events occurred in the PETRUSHKA group. Investigators judged neither event related to the tool or the prescribed antidepressant. One participant died from previously undiagnosed metastatic breast cancer, and another was hospitalized after an overdose described as being without suicidal intent.
The Trial Does Not End Prescribing Uncertainty
Patients and clinicians knew which group they were in, so expectations could have affected treatment continuation and self-reported symptoms. Outcome assessors and statisticians were intended to remain unaware of allocation, yet the authors said participants might have revealed their group during assessments. They also noted a possible disappointment effect among volunteers assigned to usual care.
Follow-up data were missing for many secondary outcomes, especially later in the trial. Sensitivity analyses using imputed data did not materially change the main direction of the findings, but the authors could not rule out a smaller benefit over time. Most participants were White, and the single Canadian secondary-care site limits conclusions about specialist settings.
PETRUSHKA also lacked adequate patient-level evidence for some antidepressants. The study covered acute monotherapy and excluded people with treatment-resistant depression, pregnancy, urgent mental-health needs and several medical conditions. Its findings therefore cannot be extended to every person taking an antidepressant or to complex combination treatment.
A Useful Tool Must Stay Inside the Consultation
The strongest evidence here is narrower than the promise of an algorithm that ends trial and error. PETRUSHKA improved one practical eight-week outcome in a randomized trial, and that is a serious result. It did not read a patient's unique brain chemistry, prove that its first choice was biologically optimal or remove the need for follow-up when symptoms or adverse effects change.
The next evidence question is whether independent use across broader populations reproduces the benefit, whether the advantage lasts and whether the system is cost-effective. Transparency also matters because the authors acknowledged that the interaction between patient characteristics, preferences and final rankings can be difficult to interpret.
Clinical decision support earns trust by making uncertainty easier to manage, not by hiding it behind the label of artificial intelligence. PETRUSHKA deserves attention because the trial measured a real patient outcome. Broad adoption should still depend on replication, clearer explanations and evidence that clinicians can use the rankings without treating them as instructions.