Diagnosis of schizophrenia from EEG with deep neural networks

Abstract

This thesis analyzes the performance of recent advances in the automatic diagnosis of schizophrenia using deep learning models on EEG data. A dataset of 81 subjects containing healthy controls (HC) and schizophrenic subjects (SZ) is used. Traditional ERP classification using LDA and shrinkage LDA is performed to get a better understanding of the separability of the data and 7 different CNN models are trained to classify the same data. The results show that linear classification using LDA performs poorly on the data, while for CNNs, it's the overfitting capabilities of the models that cause them to not generalize well on unseen data. This overfitting can go unnoticed if errors in validation are made. In that case, the performance appears much better due to the training and validation sets not being independent. This thesis shows that automatic diagnosis using deep learning models is still difficult with current methods due to the high variability of EEG signals between subjects.