Graphical representation learning-based approach for automatic classification of electroencephalogram signals in depression

Highlights

• It aims to fetch facets from EEG signals using graph representation learning.

• Graph representation learning is done using Node2vec in three levels of fusion.

• ML algorithms take these features to classify the depressed and healthy person.

Abstract

Depression is a major depressive disorder characterized by persistent sadness and a sense of worthlessness, as well as a loss of interest in pleasurable activities, which leads to a variety of physical and emotional problems. It is a worldwide illness that affects millions of people and should be detected at an early stage to prevent negative effects on an individual's life. Electroencephalogram (EEG) is a non-invasive technique for detecting depression that analyses brain signals to determine the current mental state of depressed subjects. In this study, we propose a method for automatic feature extraction to detect depression by first constructing a graph from the dataset where the nodes represent the subjects in the dataset and where the edge weights obtained using the Euclidean distance reflect the relationship between them. The Node2vec algorithmic framework is then used to compute feature representations for nodes in a graph in the form of node embeddings ensuring that similar nodes in the graph remain near in the embedding. These node embeddings act as useful features which can be directly used by classification algorithms to determine whether a subject is depressed thus reducing the effort required for manual handcrafted feature extraction. To combine the features collected from the multiple channels of the EEG data, the method proposes three types of fusion methods: graph-level fusion, feature-level fusion, and decision-level fusion. The proposed method is tested on three publicly available datasets with 3, 20, and 128 channels, respectively, and compared to five state-of-the-art methods. The results show that the proposed method detects depression effectively with a peak accuracy of 0.933 in decision-level fusion, which is the highest among the state-of-the-art methods.

Introduction

Depressive disorders are characterized by sadness, loss of interest or pleasure, disturbed sleep or appetite, a sense of guilt and hopelessness, poor concentration, and can even lead to suicide in its most severe form. According to the World Health Organization, 322 million people worldwide suffer from depression [1]. Thus, diagnosing at an early stage is critical to protect patients from the severe and irreversible consequences of depression. It has three levels, namely mild, moderate, and acute. Doctors diagnose the same based on classification criteria such as the Diagnostic and Statistical Manual of Mental Disorders Fifth Edition, Beck Depression Inventory, which are a questionnaire containing a set of questions score of which determines the level of depression. Also, the interactive sessions between the patients and health practitioners play an imperative role in detecting depression. However, these sessions may not always produce the desired results because it is dependent on the practitioner's knowledge and skill in dealing with depressive patients. Moreover, patients suffering from a mental disorder are hesitant to admit that they require treatment because they are afraid of being labeled mentally ill. As a result, clinical treatment is ineffective in assisting depressive patients in receiving proper treatment, resulting in further deterioration of the illness and its detrimental effect on an individual's life. Alternative approaches to effectively detect and diagnose depression are required to overcome the drawbacks of clinical methods. The advancement of sensor technology and communication systems has resulted in the development of electronic devices that are critical in monitoring an individual's health conditions. These devices are commonly employed to investigate the activity of the human brain in techniques such as electroencephalography (EEG), magnetoencephalography, magnetic resonance imaging, and functional magnetic resonance imaging. Out of these, EEG is a portable and non-invasive technique that uses electrodes attached to the scalp to evaluate the electrical activity of the brain. This electrical activity is represented on EEG recordings by a wavy line, which is then used by physicians and scientists to study brain functions and diagnose neurological disorders. The electrodes employed are generally useable and are labeled as F, P, O, and T, which represent the areas of the brain from which the signals are acquired, such as the frontal, parietal, occipital, and temporal lobes. The electrode in the midline of the brain is designated as z.

EEG evaluates the functioning of the brain and is used for diagnosis of illness like epileptic seizure [2], Parkinson's Disease [3], Schizophrenia [4], Stress Detection [5] and Sleep Disorder [6]. The use of EEG signals in the detection of depression is a new area of study. It begins with the extraction of EEG signals using an electroencephalogram device, followed by pre-processing, which removes disturbances in EEG signals caused by environmental changes, commonly referred to as artifacts. Following the removal of artifacts, the important features of EEG signals are extracted and fed into the classification algorithms, that determine whether or not the subject is depressed. Because of nonlinear properties EEG data, extracting useful features from it is difficult. The features and behaviour of EEG signals cannot be clearly explained using time, frequency, or time-frequency analysis. As a result, it's critical to locate stable features and create accurate models with high classification performance [7]. Hence the feature extraction procedure entails the manual extraction of handcrafted features, which is tedious and labor-intensive. The proposed method is based on an automated feature extraction process based on graphical representations, which overcomes the problem of manual feature extraction.

EEG processing entails signal processing and analysis, which begins with the extraction of useful features in the time, frequency, and time-frequency domains. This process takes a long time and a lot of human effort; additionally, there are no fixed global biomarker features for detecting depression. As a result, the work aims to reduce the manual effort required for feature extraction and propose an automatic method for detecting depressive patients using graphical representation learning. The contributions of this work are as follows:

• The proposed method aims to develop a predictive model that automatically extracts features from EEG signals using graph representation learning followed by the classification of the depressed and healthy subjects using machine learning algorithms.

• The proposed method develops three novel approaches graph-level, feature-level, and decision-level data fusion techniques that provides a way to combine the features extracted from each channel of the dataset that are obtained after applying the Node2cec algorithm in order to obtain a greater predictive power. As far as these fusion techniques are concerned, they are mainly exploiting the Node2vec algorithmic framework to carry out the data fusion procedure.

• The Node2vec algorithmic framework is adopted that computes a vector representation of a node based on random walks in the graph. These vectors corresponding to each node acts as features and is fed into machine learning algorithms to classify depressed and healthy subjects.

• The proposed method is executed on three datasets consisting of 3 channels, 20 channels, and 128 channels respectively. It is then compared with five state-of-the-art methods using different evaluation metrics, namely accuracy, sensitivity/recall, specificity, precision, and ROC curve. Empirical results illustrate that the decision-level-based proposed method outranks some existing approaches.

The paper's structure is as follows, Section II discusses a literature review of existing depression-related works. The proposed work is described in detail in Section III, which includes a detailed explanation of the Node2vec method. The results and discussions are presented in Section IV, followed by the conclusion and future scope in Section V.