Abstract / Summary
Antidepressant prescribing in major depressive disorder is guided by evidence-based recommendations, but real-world decisions often rely on clinical judgment, previous experience, and case-based reasoning. This study examined whether psychiatrists’ routine prescribing decisions could be used to identify practice-grounded patterns supporting precision-oriented treatment selection. We analyzed 428 real-world clinical cases of major depressive disorder collected from psychiatrists across several Italian centers. Thirty-one socio-demographic and clinical variables were used as input features, while prescriptions for 16 antidepressants were the output variables. Five supervised machine-learning models, including logistic regression, multilayer perceptron, decision tree, LightGBM, XGBoost, were compared with a similarity-based approach to retrieve clinically comparable cases. Across supervised models, predictive performance was limited, mostly close to random chance. In contrast, the similarity-based approach showed appreciable concordance across comparable cases. At the individual-antidepressant level, full concordance emerged in 223 cases (52.84%), while no concordance within the 10 nearest neighbors was observed in 104 cases (24.64%), increasing to 318 cases (75.36%) for full concordance at a pharmacological-class level. The limited performance of supervised models should not be viewed as algorithmic failure, but as evidence of a mismatch between conventional predictive learning and the nature of real-world psychiatric prescribing data. Conversely, similarity-based models may provide a more conceptually coherent approach, since it is closer to psychiatric analogy-based reasoning, in which clinicians often draw on previous similar cases rather than applying deterministic rules. The higher pharmacological class concordance may reflect guidelines that usually frame recommendations by therapeutic class, while preserving flexibility in the selection of individual agents.