Epilepsy Detection Using Deep Neural Networks and Explainable Artificial Intelligence Techniques

Abstract

Epilepsy is a neurological disorder in which a person experiences recurrent and unprovoked seizures. Around 50 million people have epilepsy globally. The number of epilepsy patients is increasing, as every year, 5 million new patients are confirmed to have epilepsy globally. Epilepsy is prevalent in low and middle-income countries where medical facilities are limitedly available. If epilepsy is not diagnosed at an early stage, then patients may have accidents or injuries during seizure incidents. Automated epilepsy detection systems play a significant role in the early diagnosis of epilepsy. The automated epilepsy detection system has several advantages over manual epilepsy detection. Automated systems can monitor longer Electroencephalogram (EEG) recordings without subjectivity, diagnose efficiently and accurately, and are more suitable for real-time seizure onset detection. Moreover, automated systems are useful in remote areas where trained medical experts are not available or available remotely. Such systems can be developed using advanced Machine Learning (ML) algorithms of Artificial Intelligence (AI) and Explainable Artificial Intelligence (XAI) methods. XAI methods play a significant role in explaining the predictions of AI-based epilepsy detection systems. The use of XAI methods helps these systems adapt in the medical domain for disease diagnosis. This research aims to develop an automated and interpretable epilepsy detection system and boost the classification accuracy of the system with improved preprocessing techniques. Identify the most suitable among the widely used XAI methods and Deep Neural Networks (DNNs) for epilepsy detection. Five methods are implemented to achieve these research goals. In the first method, a Bidirectional Long Short-Term Memory (Bi-LSTM) network is used to identify epileptic signals, and the XAI method, Layer-wise Relevance Propagation (LRP), is implemented in the Bi-LSTM network to explain the network predictions. The model classified epileptic and non-epileptic EEG signals with an accuracy of 87.25%. In the second method, EEG signals are converted into images by applying the Short-time Fourier Transform (STFT) to improve the classification accuracy. A DNN, suitable for image classification, i.e., the VGG16 network, is used to classify EEG signal images and achieves a classification accuracy of 99.89%. In the third method, the performance of widely used XAI methods, i.e., Shapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME) and Guided Backpropagation, is compared to explain the network predictions. It is found that the XAI method SHAP is producing the best explanations for epileptic signal identification. Therefore, the SHAP method is selected in the fourth method. The performance of four DNNs - VGG16, Resnet18, Densenet121 and Squeeznet1_0 - is compared for epilepsy detection and explanation using the SHAP method. These four networks are reported to have produced promising results for epileptic signal image classification in the literature. Therefore, these four networks are used to classify EEG signals as seizure, interictal and non-seizure. It is found that Resnet18 achieved the highest classification accuracy of 99.7% and VGG16 produced the best explanations using the XAI method SHAP. DNNs are trained using transfer learning with a fine-tuning approach. This approach takes advantage of the previous learning of the model, and the model can be trained on a small amount of data and in less time. All four methods are tested on the single-channel Bonn University dataset. The research is strengthened using a multichannel epilepsy dataset from Temple University Hospital (TUH) in the fifth method. EEG signals from the TUH dataset are converted into images by applying STFT, and their visual quality is improved by applying filters. Performance of VGG16, Resnet18 and Densenet121 is compared on EEG signal images for epilepsy detection and explanation using the SHAP method. It is found that VGG16 achieved the best classification accuracy of 97.55% using the test dataset and 97.1% ± 0.2522 using the validation dataset. Moreover, it is observed that Densenet121 produced the best results with the SHAP method. Therefore, the SHAP method is used with Densenet121 to identify the top contributing channels in the identification of epileptic patients. These channels are attached to the specific head position according to the international 10-20 system. For epilepsy patients, the top contributing channels contain seizure patterns and are attached to the brain regions from where epileptic signals originate. Thus, these channels can help neurologists identify epileptogenic brain regions and plan appropriate treatment. The XAI methods LRP and SHAP provided explanations for the deep models’ predictions and made them interpretable. This helps increase trust in the AI-based models’ decisions and adoption in the clinical decision-making process. Such a system could be beneficial for several purposes, such as initial monitoring of EEG recordings and cross-verification of epilepsy diagnosis by neurologists. Moreover, EEG technologists can use such a system to analyse EEG signals while preparing reports.