Epilepsy Detection using Bi-LSTM with Explainable Artificial Intelligence

Abstract

Neurological disorder, epilepsy, may represent as abnormal brain activities, causing seizures. Such conditions can be monitored through disturbances in normal Electroencephalogram (EEG) pattern. In literature, researchers have utilised EEG changes in epilepsy patients for automatic seizure detection using machine learning (ML). Particularly, deep learning methods noted promising accuracy in distinguishing normal and abnormal signals caused by epilepsy to classify epileptic and non-epileptic patients. However, explanation of such accuracy and decision steps are warranted for their technical verifications and clinical validations for a potential use in clinics. Explainable artificial intelligence (XAI) plays a key role in explaining such results. In this paper, we proposed a system which uses Bi-LSTM network for classification of normal and abnormal signals caused by epilepsy, and XAI method Layer-wise Relevance Propagation (LRP) to explain the predictions of the network. We implemented LRP in the Bi-LSTM network. LRP method generated relevance vector for test input vector. We reported these relevance values which indicate contribution of each data point of a signal in classification of a signal in particular class. Moreover, we displayed part of input signal, which plays a major role in signal classification, with the classification results. This is the first to use XAI method LRP with Bi-LSTM network to justify the network prediction for seizure detection in epilepsy patients. Such use of XAI method could help neurologists to build up confidence for an informed decision while using deep learning technologies in epilepsy diagnosis.

Topics & Concepts