Structural imaging predictors of ketamine response in treatment-resistant depression: a machine learning approach.

A machine-learning model using structural brain scans predicted which adults with treatment-resistant depression would respond to a single ketamine infusion. The model, trained on 99 participants, achieved 72% balanced accuracy in the discovery sample and 60% in two independent groups, with performance dropping to chance in a saline-treated control group. Greater gray matter volume in frontal regions predicted response, while greater cerebellar volume predicted non-response. The findings suggest that pre-treatment brain structure may help guide personalized treatment decisions for ketamine therapy.

Study at a glance

Abstract

Ketamine has demonstrated rapid antidepressant efficacy in treatment-resistant depression (TRD), but clinical decision-making is challenging due to variability in individual response. Current trial-and-error prescribing practices may expose patients to ineffective treatment and avoidable adverse effects, underscoring the need for reliable predictive tools to optimize treatment selection and support personalized, evidence-based care. We developed a machine-learning model (support vector classifier) to predict antidepressant response to ketamine using pre-treatment structural MRI data. The model was trained on 99 adults with TRD given a single intravenous ketamine infusion (0.5 mg/kg). Clinical response was defined as a ≥50% reduction in MADRS scores 24 h post-infusion. Internal validation used repeated nested cross-validation, and generalizability was tested in two independent ketamine-treated cohorts (n = 51) and a saline-treated control group (n = 49). Among ketamine-treated participants, 52 (52.5%) responded to treatment. The model achieved a balanced accuracy of 72.2% (sensitivity = 72.3%, specificity = 73.1%, AUC = 0.72) in the discovery sample and 60.0% (p = 0.01, AUC = 0.65) in external validation. Greater gray matter volume in frontal regions predicted response, whereas greater cerebellar volume predicted non-response. Performance dropped to chance in the saline cohort (BAC = 41.1%, AUC = 0.45), supporting pharmacologic specificity. These findings present the first machine-learning model for the prediction of ketamine response in TRD using structural neuroimaging and highlight its potential utility for stratified treatment planning and biomarker-informed interventions while providing mechanistic insight into neuroanatomical predictors of antidepressant response.