Explainable Diagnosis of Migraine via Deep Learning Through the Use of EEG Data

Migraines are a highly prevalent and costly disorder which is challenging to diagnose and typically requires a specialist reviewing a patient’s history. As a result, migraines remain underdiagnosed and hence undertreated. Electroencephalography (EEG) data has previously been used to diagnose various neurological disorders such as epilepsy, motivating the use of this data to develop a model for the automated diagnosis of migraines. In this paper, we propose a straightforward approach to automated migraine diagnosis via the fine-tuning of the ResNet-50 architecture on spectrograms of EEG data. We demonstrate that our proposed model has comparable performance to recent methods of automated migraine diagnosis at 96.3% accuracy. Furthermore, we show that we can apply methods in model explainability to highlight aspects of EEG data which our model places more importance on, making it more suitable for clinical use where the explainability of model predictions play an important factor in clinical adoption.