Development of Predictive Algorithms for Electrical Stimulation Based Major Depressive Disorder Therapy

Abstract

Major depressive disorder (MDD) is a brain disorder that is characterised by negative thoughts, mood and behaviour. Transcranial direct current stimulation (tDCS) has emerged recently as a promising brain-stimulation treatment for MDD. A standard tDCS treatment involves numerous sessions that are run over a few weeks, however, not all participants respond to this type of treatment. This delay could have negative impact upon patients that do not respond, being an inefficient use of staff time and exposing patients to ineffective treatment. The early identification of patients who respond to this type of treatment is needed. Electroencephalography (EEG) signal is a significant tool that can be used to study the modulatory effects of tDCS treatment. The significant part of this research aims to predict the clinical outcomes of tDCS treatment by analysing the patients’ EEG signals. EEG signal is a complex signal that has high sensitivity to noise.

EEG signal records both neural and non-neural activities from a large number of electrodes, therefore, the analysis and classification of EEG signals proved to be quite challenging. Machine learning has attracted the attention of many researches as a powerful approach for analysing various types of signals, including EEG. Algorithms for channel/feature selection, classification, detection, prediction and fusion have been developed. Recently, deep neural networks, particularly deep belief networks (DBNs), have emerged as a new hierarchical technique for modelling high level abstractions of data, and have been successfully applied to a number of classification problems. Similar to most classification algorithms, the existing DBNs have been mainly designed to handle single stream data, and there are hardly any attempts to generalize those to suit multi-channel signals.

Accordingly, the first part of this research investigates the utilisation of DBN to differentiate between tDCS sessions based on classification EEG signals, particularly the implementation of multi-channel DBNs. One of the important attributes that needs to be carefully studied is considering the multi-channel nature of EEG in the design and training of deep networks.

A second part of this research aims to predict which patients improve in mood and cognitive in response to tDCS treatment based on EEG data that were collected at the start of tDCS treatment. Classifying power spectral density (PSD) of resting-state EEG is achieved using support vector machine (SVM), linear discriminate analysis (LDA) and extreme learning machine (ELM). Participants were labelled as improved/not improved based on the change in mood and cognitive scores. The obtained classification results of all channel pair combinations are used to identify the most relevant brain regions. The frontal area is found to be particularly informative for the prediction of the clinical outcome of the tDCS treatment. Subject independent results reveal that our proposed method enables the correct identification of the treatment outcome for seven of the ten participants for mood improvement and nine of ten participants for cognitive improvement. This represents an encouraging sign that EEG-based classification may help to tailor the selection of patients for treatment with tDCS brain stimulation.