Randomized Trial of AI-Guided Antidepressant Selection Improving Remission and Tolerability in Primary-Care Depression

Abstract

Background: Depressive disorders represent a leading cause of disability in primary-care settings globally, yet conventional antidepressant prescribing relies predominantly on trial-and-error approaches that delay remission and increase adverse-effect exposure. Artificial intelligence (AI)-based decision-support systems offer a mechanism for personalising medication selection by integrating patient-specific clinical data, but randomised evidence from real-world primary-care settings, particularly in low- and middle-income countries, remains absent. Objective: To compare AI-guided antidepressant selection with usual clinician-directed care on remission rates, adverse-effect burden, and time-to-clinical response among adults with depressive disorders in primary care. Methods: A prospective, parallel-group randomised controlled trial was conducted in primary-care clinics across South Punjab, Pakistan. Eighty-four adults (aged 18–65 years) with PHQ-9 scores ≥ 10 initiating or switching antidepressant therapy were randomised 1:1 to AI-guided medication selection (n = 42) or usual care (n = 42). The AI tool integrated baseline symptom profiles, comorbidity data, and prior treatment history to generate individualised first-line antidepressant recommendations. Depressive severity was assessed using the PHQ-9 at baseline and at weeks 2, 4, 6, and 8; remission was defined as PHQ-9 ≤ 4; clinical response as ≥ 50% PHQ-9 reduction from baseline; and adverse effects were measured using the FIBSER scale. Outcomes were analysed using independent-samples t-tests, chi-square tests, repeated-measures ANOVA, and effect sizes reported as Cohen's d. Results: At week eight, mean PHQ-9 score was significantly lower in the AI-guided group (6.16 ± 2.4 vs 8.47 ± 2.9; mean difference 2.31, 95% CI: 1.23–3.39; p < 0.001; d = 0.88). Remission was achieved by 61.9% of AI-guided participants versus 38.1% in usual care (RR = 1.63, 95% CI: 1.04–2.54; p = 0.029). Mean time to clinical response was 3.14 ± 0.9 versus 4.86 ± 1.1 weeks (difference 1.72 weeks, 95% CI: 1.29–2.15; p < 0.001; d = 1.73). Adverse-effect burden was significantly lower in the AI-guided arm (FIBSER: 3.16 ± 1.2 vs 4.80 ± 1.4; p < 0.001; d = 1.25). No serious adverse events were recorded in either group. Conclusion: AI-guided antidepressant selection improved remission rates reduced adverse-effect burden, and accelerated clinical response compared with usual care, with consistently large effect sizes across all outcome domains. These findings support integration of AI-based decision-support tools to enhance personalised depression treatment in resource-limited primary-care settings.

Ähnliche Arbeiten