Towards Human Behavior Prediction using EEG: Analyzing Cognitive Workload and Emotional Responses

Abstract

Electroencephalography (EEG) offers a non-invasive window into brain activity, holding immense potential for predicting human behavior across diverse fields like healthcare, transportation, and Human-Computer Interaction (HCI). However, realizing this potential is hampered by several key challenges. These include the inherent complexity of EEG signals, characterized by noise and inter-individual variability, the need for sophisticated analytical methods capable of capturing complex brain dynamics, and the scarcity of large, high-quality datasets necessary for reliable analysis and model training. This research is motivated by the transformative potential of EEG and the need to overcome these challenges. Specifically, traditional analysis methods are often time-consuming, computationally expensive, and limited in their ability to fully capture the intricate dynamics of brain activity. This thesis addresses these limitations by leveraging advanced Deep Learning (DL) techniques to develop novel, scalable, and efficient methodologies for accurate and reliable EEG-based behavior prediction, presenting significant advancements in data analysis, model development, and data availability. Through three core contributions, this research offers novel methodologies and practical solutions: First, this research explores the transformative potential of transformer networks for classifying raw EEG data, obviating manual feature extraction by directly processing raw signals and capturing long-range temporal dependencies. Evaluated across age, gender, and mental workload classification, the transformer network achieved state-of-the-art performance, outperforming methods using hand-crafted features and other attention-based models. Specifically, using resting-state EEG, accuracies were 94.53% for gender and 87.79% for age. In a multitasking experiment, accuracies were 95.28% (binary) and 88.72% (ternary work load classification). This streamlined approach reduces preprocessing and domain expertise requirements, providing a strong foundation for integrating transformer models into EEG research requiring detailed temporal and spatial analysis. Second, this thesis introduces DrowzEE-G-Mamba, a DL model for robust and efficient driver drowsiness detection. This model integrates EEG data with State-Space Models (SSMs), for efficient training and inference. Key innovations include channel and depth attention modules, and the 2D-Selective-Scan (SS2D) mechanism to adapt SSMs for EEG direction-sensitivity. Evaluated on the SEED-VIG dataset, DrowzEE-G-Mamba achieved 83.24% peak accuracy, demonstrating its effectiveness in capturing local and long-range EEG dependencies. This offers improved computational efficiency with direct implications for real-time driver drowsiness detection and broader implications for cognitive state detection and Brain-Computer Interface (BCI) applications.