Individual Prediction of Remission Based on Clinical Features Following Electroconvulsive Therapy: A Machine Learning Approach

Objective: Previous prediction models for electroconvulsive therapy (ECT) responses have predominantly been based on neuroimaging data, which has precluded widespread application for severe cases in real-world clinical settings. The aims of this study were (1) to build a clinically useful prediction model for ECT remission based solely on clinical information and (2) to identify influential features in the prediction model. Methods: We conducted a retrospective chart review to collect data (registered between April 2012 and March 2019) from individuals with depression (unipolar major depressive disorder or bipolar disorder) diagnosed via DSM-IV-TR criteria who received ECT at Keio University Hospital. Clinical characteristics were used as candidate features. A light gradient boosting machine was used for prediction, and 5-fold cross-validation was performed to validate our prediction model. Results: In total, 177 patients with depression underwent ECT during the study period. The remission rate was 63%. Our model predicted individual patient outcomes with 71% accuracy (sensitivity, 86%; specificity, 46%). A shorter duration of the current episodes, lower baseline severity, higher dose of antidepressant medications before ECT, and lower body mass index were identified as important features for predicting remission following ECT. Conclusions: We developed a prediction model for ECT remission based solely on clinical information. Our prediction model demonstrated accuracy comparable to that in previous reports. Our model suggests that introducing ECT earlier in the treatment course may contribute to improvements in clinical outcomes.