OBJECTIVE: To identify important clinical or imaging features predictive of an individual's response to electroconvulsive therapy (ECT) by utilizing a machine learning approach. METHODS: Twenty-seven depressed patients who received ECT were recruited. Clinical demographics and pretreatment structural magnetic resonance imaging (MRI) data were used as candidate features to build models to predict remission and post-ECT Hamilton Depression Rating Scale scores. Support vector machine and support vector regression with elastic-net regularization were used to build models using (i) only clinical features, (ii) only MRI features, and (iii) both clinical and MRI features. Consistently selected features across all individuals were identified through leave-one-out cross-validation. RESULTS: Compared with models that include only clinical variables, the models including MRI data improved the prediction of ECT remission: the prediction accuracy improved from 70% to 93%. Features selected consistently across all individuals included volumes in the gyrus rectus, the right anterior lateral temporal lobe, the cuneus, and the third ventricle, as well as 2 clinical features: psychotic features and family history of mood disorder. CONCLUSIONS: Pretreatment structural MRI data improved the individual predictive accuracy of ECT remission, and only a small subset of features was important for prediction.
OBJECTIVE: To identify important clinical or imaging features predictive of an individual's response to electroconvulsive therapy (ECT) by utilizing a machine learning approach. METHODS: Twenty-seven depressedpatients who received ECT were recruited. Clinical demographics and pretreatment structural magnetic resonance imaging (MRI) data were used as candidate features to build models to predict remission and post-ECT Hamilton Depression Rating Scale scores. Support vector machine and support vector regression with elastic-net regularization were used to build models using (i) only clinical features, (ii) only MRI features, and (iii) both clinical and MRI features. Consistently selected features across all individuals were identified through leave-one-out cross-validation. RESULTS: Compared with models that include only clinical variables, the models including MRI data improved the prediction of ECT remission: the prediction accuracy improved from 70% to 93%. Features selected consistently across all individuals included volumes in the gyrus rectus, the right anterior lateral temporal lobe, the cuneus, and the third ventricle, as well as 2 clinical features: psychotic features and family history of mood disorder. CONCLUSIONS: Pretreatment structural MRI data improved the individual predictive accuracy of ECT remission, and only a small subset of features was important for prediction.