A Machine Learning Approach to the Smartwatch-based Epileptic Seizure Detection System

Abstract

Worldwide around 70 million people have epilepsy, and every year, more than 1 out of 1000 cases of epilepsy result in Sudden Unexpected Death in Epilepsy (SUDEP). Video – EEG is the standard clinical method for monitoring epilepsy and seizures. However, wearable systems are required to monitor epileptic activity in daily living due to the complexity of using EEG outside the laboratory. Also, to prevent SUDEP, early prediction of seizure onset is required. In this work, we propose a machine learning model to detect ictal and preictal conditions using an Empatica E4 smartwatch. The Empatica E4 records real-time photoplethysmography, electrodermal activity, accelerometry, and temperature. Clinical data were recorded from 11 patients with epilepsy (PWE) for 19 seizure onsets. Features from all the modalities were extracted by taking segments of the signal during the seizure (ictal), pre-seizure, and inter-ictal (non-seizure) conditions. These features were used on support vector machine (SVM-RBF), decision tree (DTC), and logistic regression (LRC)-based supervised training for ictal vs. non-ictal and pre-ictal vs. inter-ictal conditions. The highest accuracy of 99.40% was recorded for DTC-based seizure detection classifier during 10-fold cross-validation. Also, the highest accuracy of 95.42% was recorded for DTC-based pre-seizure onset detection classifier during 10-fold cross-validation.