Deep Learning for Multi-Class Classification of Harmful Brain Activity Using EEG Spectrogram Analysis

A study released by The Lancet Neurology shows that more than 3 million people worldwide were suffering from a neurological condition in 2021. Electroencephalography is an important tool in diagnosing and monitoring harmful brain activities, such as seizures in critically ill patients. However, manual interpretation by specialized neurologists is labor-intensive, costly, and more susceptible to variability. The present work presents the detection of six major patterns viz. – Seizure, Lateralized Periodic Discharges (LPD), Generalized Periodic Discharges (GPD), Lateralized Rhythmic Delta Activity (LRDA), Generalized Rhythmic Delta Activity (GRDA), using EEG signals. The work makes use of HMS Dataset to train a CNN-LSTM model and a ResNet50; wherein the ResNet50 model showcases comparatively better performance with higher class-specific AUPR scores: in particular, Seizure 0.7810; GPD 0.7993 scores and the weighted F1 score is 0.6921. Thus, the system can prove to be very effective at detecting these patterns from the brain activity.