Spatio-temporal features based deep learning model for depression detection using two electrodes

Abstract

Deep learning has made significant contributions to the medical field and has shown great potential in various applications. Its ability to process vast amounts of data and extraction of patterns has enabled breakthroughs in medical research, diagnosis and treatment. The application of deep learning plays a vital role in depression detection. Depression is a neurological disorder characterized by persistent feelings of sadness, hopelessness and a lack of interest. The prevalence of depression is a significant factor contributing to the rise in suicide cases on a global scale. The electroencephalogram (EEG) is a non-invasive technique used to detect depression. It records brain activity using multiple electrodes. The number of EEG electrodes used for measurement directly affects the instrumentation and measurement complexity of the experiment. The present manuscript proposes a deep learning model for depression detection, focusing on two electrodes named FP1 and FP2. The purpose of employing two electrodes is to enhance the system’s portability while reducing data acquisition time and system cost. EEG is spatio-temporal data and possesses inherent spatial and temporal features. The present manuscript proposes a methodology for extracting temporal and spatial features. The temporal feature extraction module extracts temporal features in the time domain and the spatial module extracts spatial features in the spatial domain. This manuscript presents a study on the applicability of two electrodes for depression detection. This research can enhance accessibility, user-friendliness and easier data collection and analysis. The proposed deep learning model is evaluated on two benchmark datasets. It achieves 93.41% classification accuracy, 92.54% precision, 93.23% recall, 93.06% F1 score and 97.80% area under the curve (AUC) for Hospital University Sains Malaysia dataset and for Multi-modal Open Dataset for Mental-disorder Analysis dataset it achieves 79.40% accuracy, 81.18% precision, 67.73% recall, 73.80% F1 score and 85.66% AUC.