Reliable Electrocardiogram-based Seizure Detection using Multi-Modal Electroencephalography-Electrocardiogram Representation Learning with Intra-Triplet Loss

Abstract

As more people are suffering from epilepsy and seizures by the year, ways to cure or detect these sudden seizure attacks are becoming a vital issue. Current methods of medical treatment or seizure detection devices are not practical, though: the current epilepsy treatment is very costly and the most direct way of detecting seizures, electroencephalography–the method of detecting electrical signals from the brain–is also expensive and inconvenient. Thus, I propose to use an alternate method, utilizing a seizure-detection algorithm that utilizes electrocardiography–which detects the electrical signals from the heart–trained by EEG seizure-detection algorithm as EEG is relatively much simpler to detect patterns before sudden seizures from occurring. I trained the algorithms by utilizing previously discovered seizure-detecting convolutional neural networks and by testing three different loss functions to discover the best performing method. In the final analysis, the DenseNet-169 algorithm performed the best with 90.53% accuracy utilizing the inter squared and intra squared loss functions together.