Deep learning based treatment remission prediction to transcranial direct current stimulation in bipolar depression using EEG power spectral density

Abstract

AbstractBipolar Depression (BD) characterised by changes in mood and activity levels is a leading cause of disability worldwide. Current treatments show limited efficacy. Transcranial direct current stimulation (tDCS) is a non-invasive brain stimulation method that is a potential treatment for bipolar depression. We sought to investigate the application of deep learning methods to electroencephalogram (EEG) signals to predict clinical remission. Resting-state EEG data were acquired in 21 BD participants (mean age 51.38 + 10.59 years) at baseline, using a portable 4 electrode EEG device (AF7, AF8, TP9, TP10). Treatment was 6 weeks of home-based tDCS sessions, consisting of 5 sessions per week for 3 weeks and 3 sessions per week for 3 weeks. tDCS was provided in bifrontal montage, anode over left dorsolateral prefrontal cortex (DLPFC) and cathode over right DLPFC, 2 mA, 30 minutes per session. Remission was defined as Montgomery-Åsberg Depression Rating Scale score of less than 8. Power spectral density was derived from EEG signals. Deep learning methods: 1D convolutional neural networks (1DCNNs), long short-term memory (LSTM), gated recurrent units (GRU) and their hybrid models, were investigated for prediction of remission and non-remission status following treatment. Hybrid 1DCNN and GRU model using a combination of delta, theta, and gamma band PSD from AF7 and TP10 electrodes achieved a treatment remission prediction accuracy of 79.55% (sensitivity 76.95%, specificity 83.02%). Compelling prediction accuracy for prediction of treatment remission to a course of tDCS in bipolar depression is achieved from deep learning analysis of resting-state EEG at baseline.