Machine learning algorithms for diagnosis of epilepsy from EEG

• Main Author: Thangavel, Prasanth
• Other Authors: Lin Zhiping
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Electrical and electronic engineering; Science::Medicine::Computer applications
• Online Access: https://hdl.handle.net/10356/156016
• Institution: Nanyang Technological University

Epilepsy is a set of chronic neurological brain diseases characterized by recurrent seizures. The Interictal Epileptiform Discharges (IEDs) play an important role in epilepsy diagnosis and management. According to the International Federation of Societies for Electroencephalography and Clinical Neurophysiology (IFSECN), IEDs are described as transients distinguishable from background activity, with a characteristic spiky shape, commonly, but not completely or universally, observed in interictal EEG recordings of epileptic people. Epilepsy affects about 70 million people worldwide. Epilepsy diagnosis based on IEDs in scalp electroencephalograms (EEGs) is laborious and often subjective. Therefore, it is necessary to build an effective IED detector and an automatic method to classify epileptic EEGs with IEDs vs. normal EEGs. There are only a few studies that have performed epileptic classification on EEG level, which has more clinical relevance than simply detecting IEDs. It is indeed important to consider the EEG holistically, as neurologists do, to come up with a final verdict regarding the entire EEG recording. Most studies in the literature consider a single dataset, while it is crucial to verify whether machine learning systems perform well on data from various centers; it is becoming clear in the AI community that machine learning systems tend to perform well on the datasets on which they have been trained, but may not always generalize robustly to data from other centers, seriously jeopardizing the practicality of machine learning in the medical domain. In this thesis, we developed a cascade of deep learning and shallow classifiers for reliable diagnosis of epilepsy based on IED-dependent and IED-independent features. First, we evaluate features that may provide reliable IED detection and EEG classification based on IED rate. Specifically, we investigate the IED detector based on a convolutional neural network (ConvNet) with different input features (temporal, spectral, and wavelet features) to extract additional information in the frequency domain that may not be so evident in the time-domain based ConvNet approach. We explore different ConvNet architectures and types, including 1D (one-dimensional) ConvNet, 2D (two-dimensional) ConvNet, and noise injection at various layers, which is less common in the context of signal processing and EEG analysis. We evaluate the EEG classification performance (epileptic EEGs with IEDs vs. normal EEGs) on five independent datasets. The 1D ConvNet with preprocessed fullfrequency EEG signal and frequency bands (delta, theta, alpha, beta) with Gaussian additive noise at the output layer achieved the best IED detection results with a false detection rate of 0.23/minute at 90% sensitivity. The EEG classification system obtained a mean EEG classification Leave-One-Institution-Out (LOIO) crossvalidation (CV) balanced accuracy (BAC) of 78.1% (area under the curve (AUC) of 0.839) and Leave-One-Subject-Out (LOSO) CV BAC of 79.5% (AUC of 0.856). Based on what we know, this study could likely be the one amongst the very few to conduct a cross-institutional assessment on a sizeable dataset. Since the proposed classification system only takes a few seconds to analyze a 30 min routine EEG, it may help in reducing the human effort required for epilepsy diagnosis. Traditionally, epilepsy diagnosis relies heavily on IEDs. However, since not all epileptic EEGs exhibit IEDs, it is essential to explore IED-independent EEG measures for epilepsy diagnosis. Consequently, in the latter part of the study, we explore the background characteristics of interictal EEG to develop a classification system (epileptic EEGs vs. normal EEGs) to diagnose epilepsy effectively. Specifically, we investigate the features based on spectral, wavelet, Stockwell, connectivity, graph metrics, univariate temporal measures (UTM), and patient-specific details (age and vigilance state) of EEGs from epileptic patients relative to normal subjects. The evaluation is performed on a sizeable cohort of routine scalp EEGs from five centers across Singapore, the USA, and India. In comparison with the current literature, we obtained superior LOSO CV AUC of 0.871 (BAC of 80.9%) with combination of 3 features (IED rate, and Daubechies and Morlet wavelets) for the classification of EEGs with IEDs vs. normal EEGs. The IED-independent feature UTM achieved a LOSO CV AUC of 0.809 (BAC of 74.4%). The inclusion of IED-independent features also helps to improve the EEG-level classification of epileptic EEGs with and without IEDs vs. normal EEGs, achieving an AUC of 0.822 (BAC of 77.6%) compared to 0.688 (BAC of 59.6%) for classification only based on the IED rate. To the best of our knowledge, we could be one of the very first to investigate automated algorithms for detecting epileptic EEGs without IEDs for the diagnosis of epilepsy. We successfully demonstrated in the current study that machine learning systems can be designed such that they perform well across a diversity of datasets.