Prediction of response to transcranial magnetic stimulation treatment for depression using EEG and machine learning

OBJECTIVE: Repetitive transcranial magnetic stimulation (rTMS) is an effective treatment for depression, but not for all patients. Accurate treatment response prediction could lower treatment burden. Research suggests machine learning trained with electroencephalography (EEG) data may predict response, but a limited range of features have been tested. We tested whether a combination of > 7000 time-series features were predictive of response in training and test and independent datasets. METHODS: Pre-treatment EEG from 188 patients with depression treated with rTMS were decomposed into five principal components (PCs). The highly comparative time-series analysis toolbox was used to extract 7304 time-series features from each participant and PC. A classification algorithm was trained to predict responders from these features separately for each PC. The classifier was applied to an independent dataset (N = 58) to test generalizability. RESULTS: Within the training and test dataset, the third PC showed above-chance classification accuracy (69.4 %, p = 0.005). The model generalized to the independent dataset with above-chance accuracy (60 %, p = 0.046). Analysis of feature-clusters suggested responders showed more high frequency relative power relative, and a more negative skew in the distribution of time-series values. CONCLUSIONS: Results suggest our methods could be used to inform treatment selection. SIGNIFICANCE: Our methods may enable better outcomes than 'one-size-fits-all' treatment approaches.