Predicting epiliptic seizures using machine learning applied to EEG data

Abstract

Epilepsy is a neurological condition that affects roughly 50 million people world wide. Al￾though the majority of seizures themselves are not life threatening, the loss of motor control and consciousness has a major impact on the safety of a patient. While all-cause mortality decreased 16.4% from 8756.34 per million to 7319.17 per million between 1999 and 2017, the death rate for epilepsy climbed 98.8% in the USA, from 5.83 per million in 1999 to 11.59 per million [1]. If there was a method for accurately predicting when a person will have a seizure, this mortality rate would drastically decrease, since the person would abstain from driving or doing any other dangerous activity during a time period that has been predicted to have a higher chance of seizure onset.

The aim of this dissertation is to use machine learning methods to analyse EEG signals from patients with epilepsy, and be able to predict with high sensitivity and specificity whether or not the person has a high chance of seizure onset during a certain time period.

Based on previous research conducted on the topic, it is evident that leveraging machine learning techniques to identify and predict seizures holds considerable promise. This ap￾proach has been widely explored and documented in the literature, showcasing its validity and potential.

One crucial aspect that stands out from the collective body of work is the significance of preprocessing and feature extraction in the overall process. Virtually all studies included in the literature review acknowledged the importance of these steps in enhancing the ef￾fectiveness and efficiency of their models. By carefully processing and extracting relevant features from the input data, researchers have consistently achieved improved performance and accuracy in seizure detection and prediction.

Moreover, among the various machine learning algorithms investigated for this dissertation, convolutional neural networks (CNNs) have emerged as the most popular choice and the literature survey reveals a strong prevalence of CNNs in seizure detection studies. However it is worth noting that no study included in the analysis directly compares the performance of a CNN to other models using the same dataset, thereby revealing a notable research gap.

Consequently, addressing this gap becomes a crucial next step in advancing the field of seizure prediction. Conducting comprehensive comparative studies that systematically evalu￾ate the performance of different machine learning algorithms, including CNNs, on identical datasets would contribute significantly to our understanding of their relative merits. Such investigations would enable researchers to make informed decisions regarding the most suitable algorithmic approaches for seizure detection and prediction, thereby potentially enhancing the overall accuracy and reliability of these models.

In this dissertation, EEG data is generated using a recurrent variational auto encoder, and the benefits of generated EEG data are discussed. A variational auto encoder is then used to com￾press EEG data to a latent space representation without losing any of the important features of the data. Next, a random forest ensemble model and an extreme gradient boost model are compared to see which model can most accurately classify EEG into two categories: ictal (seizure) or pre-ictal (non-seizure). Finally, three models are compared to find out which model can most accurately predict epileptic seizures. It was found that a combination of an LSTM and a CNN model was the best at predicting seizures, with the ability to predict forecast data 50 seconds into the future with a mean absolute error of 0.18.