Multiclass EEG Spectrogram Classification Using Spatial Channel Grouping and CNN for ICU Monitoring

The continuous recording of the Electroencephalogram (EEG) by Intensive Care Units (ICU) clinicians results in massive data volumes that are challenging to interpret and can take significant time, depending on the experience of the individual interpreting the data. In this paper, we propose a novel deep learning method of automatic EEG spectrogram classification into different classes for ICU brain monitoring. The classification process involves assigning an EEG pattern into one of the six clinically relevant classes, namely seizures, Ateralized Periodic Discharges (LPD), Generalized Periodic Discharges (GPD), Lateralized Rhythmic Delta Activity (LRDA), eneralized Rhythmic Delta Activity (GRDA), and Other. The primary innovation in the current research study includes a clinically motivated spatial electrode grouping method, which groups the EEG channels into four regions of the brain: left lateral, right lateral, left parasagittal, and right parasagittal. In this way, the information about the hemispheres is preserved, and the distinction between lateralized and generalized anomalies can be improved. A convolutional neural network (CNN) was trained on augmented spectrogram data with class weights, which helps in handling the high levels of class imbalance present in the HMS Harmful Brain Activity Classification Dataset. Our model resulted in a classification accuracy of 90.8% and a macro-F1 score of 0.86 with a high seizure detection F1 score of 0.86. Furthermore, a Grad-CAM visualization tool was used to provide a visualization for the predictions, while the Monte Carlo dropout technique was used to measure prediction uncertainty. In addition, our model had an Expected Calibration Error (ECE) of just 0.04, ensuring accurate confidence measures.