Abstract
As prevalent mental illnesses, anxiety disorders affect people's daily life seriously and influent our society profoundly. However, the current symptom-based diagnostic method cannot distinguish specific biological causes of specific disorders, which leads to inaccurate diagnoses and inefficient treatments. Since the non-invasive low-cost scalp electroencephalogram (EEG) contains the information of anxiety-related brain activities, it is a possible medium for anxious personality prediction and diagnostic biomarkers analysis.
Firstly, we propose a two-dimensional (2-D) conflict-focused CNN, which is based on McNaughton's conflict theory, to extract anxiety-related features in the conflict-containing EEG, and thus, to make accurate predictions. Moreover, the 2-D conflict-focused CNN is a generalized EEG feature extraction architecture that extracts temporal and spatial features separately. Our experimental results show that the 2-D conflict-focused CNN achieves excellent and stable prediction performance over the datasets with different temporal-spatial features. Meanwhile, it also provides better anxious personality prediction compared to the traditional psychology method.
Secondly, we propose a three-dimensional (3-D) EEG structure to embed the inherent spatial topology of the signals. Based on the representation of the locality and globality of EEG features, we propose a 3-D generalized CNN. It uses 3-D kernels to locate temporal-spatial features simultaneously and generate hierarchical features from local to global. Experimental results show that this architecture is capable of capturing temporal-spatial features with different complexities. Moreover, it achieves state-of-the-art anxious personality prediction performance.
Thirdly, we propose an EEG-based model analysis scheme. From the debugging side, we summarize the common but often neglected EEG debugging dilemma, the triple-blind problem. And we propose the Validation-Application-Exploration (VAE) solution to debug and verify the feature extraction ability of the models gradually. From the visualization side, we propose a decision-making visualization scheme based on Layer-wise Relevance Propagation (LRP) to visualize the features that significantly contribute to the predictions. Experimental results show that the analysis scheme can debug and tune a poorly-performing EEG-based model efficiently. Moreover, it accurately visualizes the decision-making EEG temporal-spatial features with different complexities, which sheds light on EEG-based research.