Segment Based Abnormality Detection in EEG Recordings

Electroencephalograms or EEGs are essential tests employed to diagnose any abnormalities in the brain waves that can be related to numerous brain disorders. These EEG recordings are ordered for a minimum length of 20 minutes and can go up to an hour or more, which makes the process of analysing it cumbersome and time consuming for medical professionals, and makes diagnosis very subjective. The aim of this paper is to aid in diagnosing EEG recordings with higher efficiency and accuracy. This is achieved by flagging segments of the recording into different classes that need special attention. This includes normal sleep wave patterns like POSTS, vertex waves and spindles, artifacts like ECG artifacts and Eyes open/close artifacts which can lead to misdiagnosis, and abnormal waveforms like spikes and slow waves. Most of the existing literature focuses on classifying the entire EEG as either abnormal or normal. However, in practice, a medical professional interprets the EEGs by looking at only particular sections of the recording. Therefore, our approach is to have individual segments classified instead. EEG signals, owing to their highly dynamic nature, are difficult for machine learning models to process and analyze effectively. To achieve this, we explore different methods of signal decomposition, feature elimination and classification, and find the best combination of these for the annotation task. Our combination of Empirical Wavelet Transform (EWT) for signal decomposition, Recursive Feature Elimination and Linear SVM model for classification, achieved an accuracy of 70% on the Temple University dataset and 90.78% on a private dataset.