Background <p>Depression is a heterogenous condition, with high societal and economic costs. Response to available treatments is varied, often trial-and-error, raising the need for personalization. Previous studies demonstrated success in predicting treatment outcomes for patients in psychiatric settings diagnosed with major depressive disorder, but have not done so in the general practice setting.</p> Methods <p>A retrospective cohort study analyzed electronic health records (EHRs) from the UK’s National Health Service, where depressive symptoms were recorded at two time points using the Patient Health Questionnaires (PHQ-9). Symptomatic changes between time 1 and time 2 were taken as the primary outcome of interest. Machine Learning and Causal Inference methods were used to define a data-driven antidepressant selection policy (choosing among three alternative SSRI medications) and evaluate it vis-a-vis routine clinical practice.</p> Results <p>In a cohort of 73,601 patients (65% female, mean age 49), evaluation of models for treatment selection suggested that: Reasonable treatment outcome prediction can be derived from EHRs (AUC 0.65), yet prediction is mostly driven by the initial PHQ-9 score (AUC 0.61 as a sole predictor). For essentially all patients with case-level depression prescribing antidepressants will reduce their symptoms by an expected 1.5 points For most patients, response to Citalopram, Fluoxetine and Sertraline is expected to be similar. Data driven selection of antidepressant type yields only a small improvement over current practice.</p> Conclusions <p>In the examined setting, treatment response to antidepressants in general, and to 3 specific SSRIs can be predicted reasonably well, yet the resultant treatment selection policy does not significantly improve upon current practice. Expanding this approach to a more diverse drug selection, and augmenting it with psychological, genetic and metabolomic data, are therefore future directions for research on personalized treatment selection.</p>