Deep learning methods for diagnosis of epilepsy from electroencephalograms

• Main Author: Peh, Wei Yan
• Other Authors: Lam Siew Kei
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Bioengineering; Engineering::Mathematics and analysis::Simulations
• Online Access: https://hdl.handle.net/10356/164669
• Institution: Nanyang Technological University

Brain diseases such as epilepsy, brain trauma, and stroke are serious neurological conditions and require treatments. These diseases can be diagnosed with an electroencephalogram (EEG), which can help review brain abnormalities such as excessive slowing or seizures. Hence, EEG is essential for neurological diagnosis. Unfortunately, reviewing an EEG is challenging, as EEG recordings can last for hours and be contaminated with artifacts. To automate the process, an automated EEG detector is necessary. In this study, we design three convolutional neural networks (CNN)-based pipelines to detect slowing, artifacts, and seizures in single-channel segment (channel-level), multi-channel segment (segment-level), and full EEG (EEG-level). First, we develop a CNN-based slowing detector to detect pathological slowing at the channel-, segment-, and EEG-level. Across four datasets, the detector reports a mean balanced accuracy (BAC) of 82.0% and 81.8% at the EEG-level for leave-one-institution-out (LOIO) and the leave-one-subject-out (LOSO) cross-validation (CV), respectively. Furthermore, the detector can process a 30-minute EEG in 4 seconds, allowing real-time deployments. Second, we design a CNN-transformer with a belief matching (BM) loss (CNN-TRF-BM) model to detect chewing, electrode pop, eye movement, muscle, and shiver artifacts at the channel- and segment-level. The artifact detector reports a sensitivity (SEN) of 0.420, 0.320, and 0.133, subject to a fixed specificity of 95%, 97%, and 99%, respectively. The artifact detector can identify and reject artifacts with high specificity, leading to a cleaner EEG for visual evaluations. Finally, we develop a patient-independent CNN-TRF-BM-based seizure detector to detect seizures at the channel-, segment-, and EEG-level. Across six datasets, the system achieves EEG-level seizure detection average sensitivity (SEN) of 0.227-1.00, average precision (PRE) of 0.377-1.00, and median false positive rate per hour (FPR/h) of 0-0.559. The seizure detector performs detection in less than 15 seconds for a 30-minute EEG and can accelerate the EEG reviewing process. In conclusion, the proposed slowing, artifact, and seizure detectors can tremendously boost the review of EEG in clinical practice.