Antidepressants are known to help many people with depression, but the same drug does not work equally well for everyone. Treatment response and side effects can vary depending on a number of factors such as age, sex, life history, or other medical conditions, which are important for clinicians to consider when prescribing for the first time.
However, the problem is that, in practice, many clinicians do not have the time, capacity, or specialist information needed to tailor prescribing decisions using this evidence. Patients also differ in which side effects they would most want to avoid, and this has an impact on whether they will continue taking them (Coe et al., 2023). This is important because antidepressants usually need to be taken consistently to benefit from treatment, but side effects or other initial adverse events lead many people to stop taking them too early (Coe et al., 2023; Tomlinson et al., 2020).
Taking all this into consideration, a recent study by Andrea Cipriani and colleagues (2026) tested whether a digital tool could make antidepressant prescribing practices more personalised, by combining predictions about treatment fit with patients’ own preferences.
Methods
The authors developed a digital tool called PETRUSHKA to help clinicians choose which antidepressants were likely to work best for individual patients. Recommendations were based on prediction models (developed using machine learning) using personal characteristics (e.g., sex, age, childhood maltreatment, comorbid medical conditions), alongside patient preferences about which adverse side effects they would most like to avoid.
They tested this tool compared to usual care (i.e., based on the normal procedure clinicians would use to prescribe antidepressants), in a randomised controlled trial (RCT) across 47 sites in Brazil, the UK and Canada. Overall, 540 participants were randomised, although 20 were subsequently found to be ineligible and excluded from the analyses. They looked at whether there were differences in the number of people who had discontinued treatment after 8 weeks between those who had been prescribed antidepressants based on the digital tool versus those in usual care, as well as changes in depressive and anxiety symptoms at 8 and 24 weeks.
Results
Sample characteristics
Of the 520 eligible participants, the primary analysis included 493 people diagnosed with major depressive disorder (MDD), with 241 randomised to PETRUSHKA and 252 randomised to usual care. Participants had a median age of 35 years, and the majority were female (58%) and white (84%). In PETRUSHKA, the most commonly prescribed antidepressants were mirtazapine (29%), escitalopram (28%) and vortioxetine (24%), whereas the most commonly prescribed antidepressants in usual care were sertraline (52%), citalopram (15%) and fluoxetine (9%).
Main findings
After 8 weeks of antidepressant treatment, 41 participants randomised to PETRUSHKA in comparison to 69 participants randomised to usual care discontinued their antidepressants, representing a reduced risk of discontinuation of almost 40% (risk ratio = 0.62, 95% CI [0.44 to 0.88], p = .007). However, there was no clear evidence of this effect at 24 weeks.
The authors also looked at whether participants stopped treatment due to an adverse event and found that participants in the PETRUSHKA group were less likely to stop antidepressants because of an adverse event compared to those in the usual care group (risk ratio = 0.59, 95% CI [0.36 to 0.97], p = .04). However, this difference was also not evident at 24 weeks. There was also some indication that the effect was larger in primary care than secondary care.
There were no clear differences between the two groups in depression or anxiety symptoms at 8 weeks. In contrast, by 24 weeks, self-reported depression and anxiety symptoms were lower in the PETRUSHKA group in comparison to the usual care group (depression: mean difference = -1.92, 95% CI [-3.06 to -0.78]; anxiety: mean difference = -1.39, 95% CI [-2.26 to -0.52]). Although these differences favoured PETRUSHKA, they were relatively small compared with the minimal clinically important changes reported for these measures (5 points for depression and 4 points for anxiety), although these thresholds refer to changes within individuals rather than differences between groups.
A similar pattern was found when they looked at depression and anxiety measured through observer ratings, although these estimates were more uncertain and not statistically significant. Finally, self-reported health-related quality of life was also higher in the PETRUSHKA group at 24 weeks (70.3 vs 66.8; p = .04), although this finding was based on substantially fewer participants than were originally randomised.
Conclusions
Overall, the trial found that individuals with major depressive disorder who were prescribed an antidepressant through the PETRUSHKA tool were less likely to stop treatment during the first 8 weeks than those prescribed through usual care. The PETRUSHKA group also showed greater improvements in self-reported depression and anxiety symptoms at 24 weeks. These findings suggest that shared decision-support systems may help to personalise antidepressant treatment by combining prediction modelling with individual patient preferences to calculate an overall recommendation score for each antidepressant. However, it remains unclear whether the benefits were driven by more accurate prediction of treatment fit, greater patient involvement (and therefore motivation to continue treatment), or a combination of the two, meaning more research is needed.
Strengths and limitations
A major strength of this study was the authors’ extensive engagement with people with lived experience, as well as international experts and ethicists, when deciding the most appropriate outcome and comparator for the study. This means that the study reflected the priorities and perspectives of people who may actually use and benefit from this type of digital tool, rather than only reflecting what the researchers had decided was important, increasing its potential utility and impact.
Another important strength was the incorporation of patient preferences within the decision-making tool. Patients’ perspectives are often underestimated in treatment decisions, but they may be particularly important when there are several treatment options available, each with different side effects. People do not judge all side effects or adverse events in the same way, so a treatment that is acceptable to one person may be unacceptable to another. This may be especially relevant in mental health care, where involvement in treatment decisions may increase a patient’s sense of agency and willingness to continue treatment (Gurtner et al., 2021).
The study was conducted across 47 sites in the UK, Brazil and Canada, improving confidence that the findings are not limited to a single setting. However, there was substantial missing data for some secondary outcomes, particularly at 24 weeks. Although the authors tried to retrieve missing data from participants, clinicians and sites, data were still incomplete. Sensitivity analyses accounting for missing data produced similar results, but this makes it difficult to untangle whether the limited effects on antidepressant discontinuation at 24 weeks reflect a true reduction in the effect, or whether they were influenced by missing data. Similarly, the beneficial effects of the PETRUSHKA tool on depression and anxiety at 24 weeks were based on fewer participants (n = 333-334 at 8 weeks versus n = 258-259 at 24 weeks), so the results may have been influenced by selective drop out, which could have affected the estimated benefits of the tool.
A further limitation is that while the authors chose a clinically meaningful primary outcome, the applicability of the trial is limited by the fact that they only considered antidepressant monotherapy and did not include other effective interventions, such as psychotherapy, when in reality a substantial proportion of people receive antidepressants alongside psychotherapy (Gaspar et al., 2019).
Finally, as the authors acknowledged, clinicians and patients knew which treatment group participants had been allocated to. This was difficult to avoid due to the nature of the intervention, but it may still have influenced how clinicians discussed treatment options or how patients judged their own symptoms. The authors also noted that participants who signed up for a trial of personalised treatment but were allocated to usual care may have been disappointed, which could have influenced their likelihood of continuing treatment. The fact that clearer effects were found for self-reported depression and anxiety symptoms may indicate some influence of treatment awareness, although this cannot definitely be concluded.
Implications for practice
This study by Cipriani et al. (2026) provides initial evidence for a more personalised approach to prescribing in psychiatry. The PETRUSHKA tool combined prediction modelling with patients’ own treatment preferences, rather than relying only on average treatment effects or clinicians’ judgements. This aligns with qualitative findings that prediction models in psychiatry should be implemented in ways that preserve human involvement and personal choice (Higson-Sweeney et al., 2026; read Ana’s Mental Elf blog). This approach could provide a useful model for other areas of mental health care where treatment selection is difficult and discontinuation is common, such as ADHD (Gajria et al., 2014) and bipolar disorder (Goldberg, 2019). Similar decision-support tools could be developed for other pharmacological treatments, and potentially also adapted to support decisions about psychological therapies, where patient preference and treatment fit may also influence engagement and outcomes (Swift et al., 2018).
The finding that the tool appeared more effective in primary care than in specialist psychiatric settings is also interesting, although this comparison should be interpreted cautiously. The authors suggest that this may be due to the specialist knowledge of psychiatrists in secondary care settings, which may lead to better tailoring of antidepressants. This suggests that digital decision support tools may be particularly useful in settings where depression is commonly treated by non-specialist clinicians, including primary care and lower resource settings where access to specialist mental health care is limited. However, given the small number of participants recruited from secondary care, differences between settings would need to be tested directly in future research.
Overall, a tool that helps patients start an antidepressant that they are more willing and able to continue could improve the chances of sustained treatment and, in turn, symptom improvement and quality of life. If these benefits are confirmed in future studies, this could help reduce the personal and healthcare burden of depression.
Statement of interests
Oonagh Coleman has no conflicting interests to declare.
Edited by
Dr Nina Higson-Sweeney.
Links
Primary paper
Andrea Cipriani, Karen Fernandes, Benoit Mulsant, Orestis Efthimiou, Nicola Williams, Sam Mort, … & Edoardo Ostinelli (2026). A decision-support system to personalize antidepressant treatment in major depressive disorder: a randomized clinical trial. JAMA, 335(14), 1219-1231. https://doi.org/10.1001/jama.2026.1327
Other references
Coe, A., Gunn, J., Fletcher, S., Murray, E., & Kaylor-Hughes, C. (2023). Self-reported reasons for reducing or stopping antidepressant medications in primary care: Thematic analysis of the diamond longitudinal study. Primary Health Care Research & Development, 24, e16. https://doi.org/10.1017/S1463423623000038
Gajria, K., Lu, M., Sikirica, V., Greven, P., Zhong, Y., Qin, P., & Xie, J. (2014). Adherence, persistence, and medication discontinuation in patients with attention-deficit/hyperactivity disorder: A systematic literature review. Neuropsychiatric Disease and Treatment, 10, 1543–1569. https://doi.org/10.2147/NDT.S65721
Gaspar, F. W., Zaidel, C. S., & Dewa, C. S. (2019). Rates and determinants of use of pharmacotherapy and psychotherapy by patients with major depressive disorder. Psychiatric Services, 70(4), 262–270. https://doi.org/10.1176/appi.ps.201800275
Goldberg, J. F. (2019). Personalized pharmacotherapy for bipolar disorder: How to tailor findings from randomized trials to individual patient-level outcomes. Focus, 17(3), 206–217. https://doi.org/10.1176/appi.focus.20190005
Gurtner, C., Schols, J. M. G. A., Lohrmann, C., Halfens, R. J. G., & Hahn, S. (2021). Conceptual understanding and applicability of shared decision-making in psychiatric care: An integrative review. Journal of Psychiatric and Mental Health Nursing, 28(4), 531–548. https://doi.org/10.1111/jpm.12712
Higson-Sweeney, N., Peycheva, A., & Breedvelt, J. J. F. (2026). ‘It is just a prediction; it’s, like, not fact’: Youth attitudes towards risk prediction tools and personalised preventive interventions for depression and anxiety. BMJ Mental Health, 29(1), e302327. https://doi.org/10.1136/bmjment-2025-302327
Swift, J. K., Callahan, J. L., Cooper, M., & Parkin, S. R. (2018). The impact of accommodating client preference in psychotherapy: A meta-analysis. Journal of Clinical Psychology, 74(11), 1924–1937. https://doi.org/10.1002/jclp.22680
Tomlinson, A., Furukawa, T. A., Efthimiou, O., Salanti, G., De Crescenzo, F., Singh, I., & Cipriani, A. (2020). Personalise antidepressant treatment for unipolar depression combining individual choices, risks and big data (PETRUSHKA): Rationale and protocol. Evidence-Based Mental Health, 23(2), 52–56. https://doi.org/10.1136/ebmental-2019-300118
Veic, A. (2026). Predicting depression and anxiety: what young people really think. The Mental Elf.