Stroke Detection using Machine Learning on EEG Data

Sammanfattning

Stroke is a devastating and common medical condition, which one in four people will experience in their lifetime. Strokes often occur suddenly and without any prior signs, and are hard to notice as they occur invisibly within the brain; only the appearance of outward symptoms give away the ongoing attack. This, combined with the fact that the damage done by the stroke worsens the longer it goes untreated, makes more affordable and reliable stroke detection methods highly desirable. The goal of our project was to examine the feasibility of detecting stroke in patients using only brain activity data collected with easily used and relatively low cost electroencephalography (EEG) machines. To accomplish this, EEG datasets from patients with and without stroke were collected, standardized, and preprocessed using different methods. The different variations of this data was then used to train machine learning models, and the ability of each trained model to identify EEG recordings of stroke patients was tested. It was found that the method of preprocessing of the data was critical for the model to achieve good results. Using certain preprocessing methods, and with a sufficient amount of data, the model was able to correctly differentiate stroke and non-stroke EEG recordings in the dataset with 100% accuracy. However, a number of issues which may have had an effect on the result were identified, including the lack of relevant publicly available datasets, the small size of the used datasets, and potential issues in the standardization process.