Epilepsy is one of the most common neurological disease in the world. Researchers focus on automatic electroencephalogram (EEG) seizure detection methods and achieve remarkable detection accuracy. However, there still exist highly uncertain decisions, which may be caused by the noise signals such as ocular artifacts. Considering the misclassification of an epileptic patient as non-seizure could have severe consequences, we point out that the automatic EEG seizure detection methods should explicitly provide the reliability of the detection. To achieve this goal, we propose an Evidential Multi-view Learning (EML) method, which reduces noises by multiple feature extraction methods and make trusted decision accordingly. EML constructs initial multi-view features to capture the diverse information of EEG signals. For each view, EML establishes deep neural networks to learn high-level features and view-specific evidence of each category, which could be termed as the amount of support to each category collected from data. By dynamically fusing different views at the evidence level, EML makes reliable prediction accordingly (strong evidence indicates high prediction reliability). We empirically evaluate EML on a public EEG Epileptic Seizure Detection dataset (CHB_MIT). Experiments show the superiority of EML over state-of-the-art baseline methods. We will release the code in https://github.com/Wednesque/Trust-EEG-Epileptic-Seizure-Detection-via-Evidential-Multi-view-Learning.
Read full abstract