Abstract
Purpose: Electroencephalogram (EEG) is a fundamental diagnostic tool for epilepsy. However, it is difficult to determine abnormal signals such as spike and wave discharges in some EEG waveforms. This study aimed to compare machine learning (ML) and physician-annotated readings for the same EEG data.
Methods: Ten abnormal interictal EEG recordings from patients with childhood epilepsy with centrotemporal spikes were selected for analysis. Five EEG recordings were examined by two medical doctors to build an ML model with a convolutional neural network. Five additional EEG recordings were manually labeled by two medical doctors, two clinical technicians familiar with EEG, and one novice university student as evaluating datasets. The statistical sensitivity, specificity, area under the receiver operating characteristic curve (AUC), accuracy, precision, and recall of the ML model for each evaluating group were obtained. The label match ratio was determined as the ratio of the number of matched labels to the number of all labels manually annotated by the annotator.
Results: The total annotation count was 37,752 (7,167 for training and 30,585 for testing). The mean label match ratio was 98.9% and the lowest was 93.0%. The sensitivity was 99%. The AUC for ML was > 0.95, and the novice student had a lower AUC with statistical significance. The label match ratio did not show a significant difference.
Conclusion: ML models are candidate training tools for EEG analysis.
