Interpretable and robust AI in electroencephalogram systems

摘要

Electroencephalogram (EEG) provides valuable information about brain activities and states in a non-invasive way, making it a crucial research area in human-computer interaction (HCI). With the rapid advancement of artificial intelligence (AI) technologies, EEG systems have increasingly harnessed the power of AI for various clinical, entertainment, and social interaction applications. For example, sleep staging systems combine EEG signals with deep learning to assist physicians in the rapid diagnosis of sleep disorders. Driver monitoring systems employ EEG-based deep neural networks (DNNs) to accurately detect driver fatigue, thereby reducing the risk of car accidents. Additionally, robotic arm control systems use DNNs to translate human thoughts, as reflected by EEG signals, into control signals, enabling disabled individuals to perform basic tasks such as drinking water or moving objects. Despite the significant progress driven by AI, models, particularly those based on deep learning, remain largely unexplainable due to their black-box nature. This lack of interpretability poses challenges in understanding and trusting the AI's decisions. Furthermore, these models are susceptible to both intentional and unintentional attacks, raising serious concerns about their robustness and reliability. Addressing these issues is crucial for AI's widespread adoption and safe deployment in EEG systems. Researchers are actively exploring methods to enhance the interpretability and robustness of AI models, ensuring they can provide reliable and transparent support in critical applications. As the field evolves, these advancements will be pivotal in realizing the full potential of AI-enhanced EEG systems for improving human life across various domains. A comprehensive literature review on interpretable and robust AI techniques for EEG systems is presented in this thesis. It begins with an introduction to the background knowledge of EEG signals. Next, it proposes a taxonomy of interpretability, categorizing it into three types: backpropagation, perturbation, and rule-based methods. Additionally, it categorizes robustness based on undesirable factors into four classes: noise and artifacts, human variability, data acquisition instability, and adversarial attacks. The literature review includes detailed analyses and comparisons for each category. Finally, it identifies several critical challenges for interpretable and robust AI in EEG systems and discusses their future directions. This literature review aims to guide researchers in understanding this field's latest advancements and future trends. This thesis introduces a novel framework called the Hybrid Attention EEG Sleep Staging (HASS) Framework, designed for cross-subject EEG sleep staging tasks. HASS employs a spatio-temporal attention mechanism to adaptively assign weights to inter-channel and intra-channel EEG segments based on the spatio-temporal relationships of the brain during different sleep stages. Experimental results on the MASS and ISRUC datasets demonstrate that HASS can significantly improve typical sleep staging networks. HASS addresses the difficulties of capturing the spatial-temporal relationships of EEG signals during sleep staging under cross-subject scenarios and holds promise for improving the accuracy and reliability of sleep assessment in both clinical and research settings. Furthermore, this thesis introduces EENED and GlepNet, novel EEG-based architectures for neural epilepsy detection. To address the challenge of learning robust global-local representations in epilepsy signals, EENED/GlepNet combines temporal convolutional layers with a multi-head attention mechanism. This approach enhances the performance of epilepsy diagnosis, potentially improving patient outcomes. The utilization of Grad-CAM for interpretability further elevates its clinical value, allowing healthcare professionals to validate and visually understand the model's diagnostic process. This advancem