Autonomous Seizure Prediction: Using machine learning techniques to detect pre-seizure brain states

Abstract

Nearly 1% of the world’s population is afflicted by epilepsy. It is the fourth most common neurological disorder, and is characterized by unprovoked seizures, and other health issues. Arguably, the most detrimental characteristics of epilepsy are the spontaneity and irregularity of seizure attacks. Since epileptic patients do not know when a seizure may occur, they are at greater risk, or forbidden from performing dangerous activities like driving or swimming, and are often prescribed medication with harmful side effects.An accurate seizure prediction algorithm would help these patients live more regular lives. Clinical studies have shown evidence of pre-seizure brain activity patterns, and thus, a large amount of research is being performed on predicting seizures. Still, no automated seizure prediction algorithm has come close to the reliability needed for practical use.The goal of this project is to use machine learning to create that algorithm. We downloaded data from, and competed in the Melbourne University Seizure Prediction Challenge on Kaggle, and created a model to distinguish between interictal and preictal iEEG clips. The final model we create scores an AUC of 0.76715 on unseen test data.