Biosignal-Based Brain State Evaluation Using Deep Learning Frameworks

Abstract

Biosignals serve as invaluable tools for monitoring the physiological states of specific body parts, aiding in the evaluation and treatment of physical and mental conditions. Over the years, machine learning has been extensively utilized to assess and predict diverse physical and mental states based on biosignals. Traditional machine learning models require structured pipelines to convert raw signals into learnable datasets with key features based on prior expert knowledge, ensuring effective classification. On the other hand, deep learning models have overcome these limitations by enabling end-to-end learning and exhibiting superior performance in classification. In this dissertation, we interpret brain states based on biosignals, particularly electroencephalography (EEG) and electrocardiography (ECG), and propose novel methods for evaluating and predicting these states using deep learning techniques. The first part explores neural components associated with categorical speech perception using single-trial EEG and demonstrates the effective differentiation of corresponding brain activities through a bidirectional long short-term memory network (BiLSTM). The second part investigates the effects of psychiatric illnesses on biosignals and suggests deep learning architectures suitable for discriminating the severity of these conditions. First, we reveal the interrelationship between heart rate variability (HRV) and depressive or anxious symptoms, examining the clinical applicability of HRV indices in estimating their severity. Additionally, we propose a deep learning model that effectively classifies the severity of psychiatric symptoms based on time, frequency, and nonlinear domain HRV indices. We anticipate that this research will highlight the significance of exploring novel methodologies for assessing brain states and serve as a scaffold for the development of practical automated discrimination and diagnostic systems in the future.