Electroconvulsive therapy (ECT) is one of the most effective treatments in cases of severe and treatment resistant major depression. 60-80% of patients respond to ECT, but the procedure is demanding and robust prediction of ECT responses would be of great clinical value. Predictions based on neuroimaging data have recently come into focus, but still face methodological and practical limitations that are hampering the translation into clinical practice. In this retrospective study, we investigated the feasibility of ECT response prediction using structural magnetic resonance imaging (sMRI) data that was collected during ECT routine examinations. We applied machine learning techniques to predict individual treatment outcomes in a cohort of N = 71 ECT patients, N = 39 of which responded to the treatment. SMRI-based classification of ECT responders and non-responders reached an accuracy of 69% (sensitivity: 67%; specificity: 72%). Classification on additionally investigated clinical variables had no predictive power. Since dichotomisation of patients into ECT responders and non-responders is debatable due to many patients only showing a partial response, we additionally performed a post-hoc regression-based prediction analysis on continuous symptom improvements. This analysis yielded a significant relationship between true and predicted treatment outcomes and might be a promising alternative to dichotomization of patients. Based on our results, we argue that the prediction of individual ECT responses based on routine sMRI holds promise to overcome important limitations that are currently hampering the translation of such treatment biomarkers into everyday clinical practice. Finally, we discuss how the results of such predictive data analysis could best support the clinician's decision on whether a patient should be treated with ECT.
Keywords: Electroconvulsive therapy; Machine learning; Magnetic resonance imaging; Major depression; Response prediction.
Copyright © 2020 Elsevier Inc. All rights reserved.