Identification of Relevant ECG Features for Epileptic Seizure Prediction Using Interpretable Machine Learning

Epileptic seizure prediction holds the potential to enhance the quality of life for individuals with epilepsy by enabling the possibility of timely administration of medication and first aid, as well as preventing subsequent accidents. In this paper, we consider the well-established Heart Rate Variability (HRV) and Lorenz features, and augment them with the electrocardiogram (ECG) multifractality features for the first time for seizure prediction. Our experimental results demonstrate that incorporating multifractality features significantly enhances epileptic seizure prediction, with a 7.5% improvement over using only HRV features and a 6.9% improvement over using both HRV and Lorenz features. We also investigate the significance and impact of features in a seizure prediction Machine Learning (ML) model utilizing ECG signals, aiming to shed light on the intricate relationship between cardiac function and epileptic seizures. We employ the interpretability framework, SHAP (SHapley Additive exPlanations), to interpret the prediction patterns. Based on our analysis, multifractality features are among the most important features in seizure prediction, capturing patterns that are not captured by the HRV and Lorenz features.