Beyond trial-and-error: AI-guided antidepressant selection using integrated genetic and clinical data in major depressive disorder

Background

Antidepressant prescribing in major depressive disorder (MDD) remains largely empirical, relying on a trial-and-error approach that often results in delayed treatment response, increased burden of side effects, and reduced patient adherence. Despite advances in psychopharmacology, clinicians still lack robust tools to guide individualized treatment selection, contributing to prolonged illness burden and functional impairment. Emerging approaches integrating multi-dimensional patient data—including genetic, clinical, and socio-demographic variables—offer the potential to improve treatment precision. Artificial intelligence (AI)-based clinical decision support systems may provide a framework for synthesizing such complex datasets into actionable recommendations, thereby facilitating more personalized and efficient treatment strategies in routine clinical practice. Furthermore, variability in individual treatment response highlights the need for approaches that better account for biological and clinical heterogeneity across patients.

Objectives

To evaluate whether an AI-based decision support system (Predictix), integrating genetic, clinical, and socio-demographic data, can support personalized antidepressant selection and lead to clinically meaningful outcomes in patients with MDD in a real-world clinical setting.

Methods

In this single-center, open-label pilot study conducted at Pitié-Salpêtrière Hospital (Paris), 30 adult patients meeting DSM-5 criteria for MDD initiated antidepressant treatment based on recommendations generated by the Predictix system. The algorithm integrates pharmacogenetic data, clinical characteristics, and socio-demographic factors to provide ranked treatment options. The system generates structured recommendations intended to support, rather than replace, clinician decision-making in routine practice. Participants were followed over 8 weeks. The primary outcome was treatment response, defined as a ≥50% reduction in the Quick Inventory of Depressive Symptomatology–Self-Report (QIDS-SR-16). Secondary outcomes included response rates based on the Patient Health Questionnaire-9 (PHQ-9) and Clinical Global Impression–Improvement (CGI-I) scale, as well as measures of tolerability and acceptability, including the Frequency, Intensity, and Burden of Side Effects Rating (FIBSER). Analyses were conducted on study completers (n=18).

Results

Among study completers, response rates were 55.6% (QIDS-SR-16) and 61.1% (PHQ-9). According to CGI-I ratings, 72.2% of participants were classified as responders, and 94.4% were rated as clinically improved (CGI-I ≤3). Treatment was generally well tolerated, with 66.7% of participants reporting low side-effect burden (FIBSER ≤2). Overall acceptability of the AI-guided approach was favorable among participants. However, turnaround time required for genetic analysis emerged as a notable practical limitation, potentially affecting real-world scalability and implementation. No serious adverse events related to treatment selection were observed during the study period.

Conclusions

These findings provide preliminary clinical support for AI-guided antidepressant selection as a feasible and data-driven approach in the management of MDD. By integrating biological and clinical information, such systems may facilitate more informed and individualized treatment decisions in routine psychiatric practice. These findings also highlight the potential role of AI-based tools in improving treatment efficiency and reducing the burden associated with repeated treatment failures. While promising, these results should be interpreted cautiously given the pilot design and limited sample size. Further validation through larger, randomized controlled studies is required to establish efficacy, generalizability, and cost-effectiveness of AI-assisted prescribing in real-world settings.

References

No references