ETSNet: A deep neural network for EEG-based temporal–spatial pattern recognition in psychiatric disorder and emotional distress classification

Highlights

• First study to explore end-to-end solution to enhance schizophrenia classification.

• Generating whole-brain source signals from raw EEG using realistic head model.

• EEG Temporal Spatial Network (ETSNet) for source signals of regions of interest.

• Created two variants of ETSNet: single (ETSNet${}^s$), and fusion network (ETSNet${}^f$).

• The performance of ETSNet${}^s$ is evaluated on two publicly available dataset.

Abstract

The use of EEG for evaluating and diagnosing neurological abnormalities related to psychiatric diseases and identifying human emotions has been improved by deep learning advancements. This research aims to categorize individuals with schizophrenia (SZ), their biological relatives (REL), and healthy controls (HC) using resting EEG brain source signal data defined by regions of interest (ROIs). The proposed solution is a deep neural network for the cortical source signals of the ROIs, incorporating a Squeeze-and-Excitation Block and multiple CNNs designed for eyes-open and closed resting states. The model, called EEG Temporal Spatial Network (ETSNet), has two variants: ETSNet${}^s$ and ETSNet${}^f$. Two evaluations were conducted to show the effectiveness of the proposed model. The average accuracy for the classification of SZ, REL, and HC using EEG resting data was 99.57% (ETSNet${}^f$), and the average accuracy for the classification of eyes-open (EO) and eyes-closed (EC) resting states was 93.15% (ETSNet${}^s$). An ablation study was also conducted using two public datasets for intellectual and developmental disorders and emotional states, showing improved classification accuracy compared to advanced EEG classification algorithms when using ETSNet${}^s$.

Introduction

The use of neurophysiological signals, such as the electroencephalogram (EEG), has been widely adopted for the study of brain dysfunction in psychiatric disorders. Machine learning techniques have also been utilized to develop EEG-based psychiatric diagnostic systems. With approximately 450 million people suffering from psychiatric disorders worldwide, including schizophrenia, bipolar disorder, and depression [1], reliable risk prediction and diagnosis using EEG remains a challenge due to the complexity and high level of noise in EEG signals and the significant differences between healthy and mentally ill individuals.

Previous research on computer-aided diagnosis (CAD) systems for schizophrenia has focused on distinguishing individuals with schizophrenia (SZ) from healthy controls (HC). However, such systems can lead to misdiagnosis in individuals who share genetic characteristics and phenotypes (i.e., endophenotypes [2]) with SZ but do not have the disorder. To improve the accuracy of the CAD system, genetic susceptibility to schizophrenia must be taken into consideration. Including features that differentiate biological relatives of SZ (REL) from their probands can reduce the risk of misdiagnosis. Although REL shares genetic traits with SZ, they do not exhibit the disease-specific traits that are present in SZ and absent in HC [3]. As such, it is ideal to develop a CAD system for diagnosing SZ by using features derived from both REL and SZ and HC [4], [5].

Machine learning techniques, such as support vector machine (SVM) and logistic regression, have been applied to EEG signals to diagnose schizophrenia [6], [7], [8], [9], [10]. However, these studies have only utilized straightforward EEG features at the scalp sensor level, which can limit the accuracy of the diagnosis due to the spatial resolution limitations of sensor-level signals. To improve the classification accuracy, brain source-level signals with high spatial and temporal resolution should be utilized, as they provide more rich and detailed data [11].

There has been a growing interest in the use of deep learning in EEG analysis for neuropsychiatric disorders. Deep learning has the potential to provide deeper insights into EEG analysis and contribute to more effective and timely treatment of these illnesses. Recent studies utilizing deep learning models have shown significant improvements in the accuracy of classifying individuals with SZ versus HC compared to conventional machine learning methods [12], [13], [14]. Given the advancements in deep learning, it is likely that this technology will significantly enhance the CAD system for SZ diagnosis through the application of deep-learning techniques.

The aim of this work is to develop a deep learning model for the categorization of individuals with SZ, REL, and HC. The paper’s contributions are summarized as follows:

• This study is the first to explore an end-to-end solution for EEG-based classification of SZ, REL, and HC to enhance the schizophrenia diagnostic classification system, as depicted in Fig. 1.

• The focus is on obtaining high spatial and temporal resolution characteristics from EEG data by computing whole-brain cortical source signals using a realistic head model built from individual magnetic resonance imaging (MRI) data.

• A novel EEG Temporal Spatial Network (ETSNet) is proposed for the cortical source signals of regions of interest (ROIs) by incorporating a Squeeze-and-Excitation block and fusing networks generated in the eyes-open (EO) and eyes-closed (EC) resting states. Two variants of ETSNet are developed: single (ETSNet${}^s$), and fusion network (ETSNet${}^f$).

• Extensive experiments are conducted to evaluate the performance of the proposed deep neural model in classifying SZ, REL, and HC, as well as differentiating EO vs. EC resting states. ETSNet${}^f$ achieved a classification accuracy of 99.57% for the 3 classes, SZ, REL, and HC. ETSNet${}^s$ achieved an accuracy of 93.15% in distinguishing EO vs. EC resting states.

• The performance of ETSNet${}^s$ is evaluated on two publicly available datasets, SEED [15] and IDD [16]. The results show improved classification accuracy on these datasets with ETSNet${}^s$.