Abstract Photosensitivity is a neurological condition in which the brain produces epileptic reactions known as Photoparoxysmal Responses (PPR) to certain visual stimuli, which may trigger epileptic seizures in the worst cases. This pathology is diagnosed through a standardized protocol called Intermittent Photic Stimulation consisting of stimulating the patient with a flashing light while recording the brain activity by Electroencephalogram (EEG) until a PPR is triggered. Due to the nature of the stimulation process and the photosensitivity’s low prevalence, the automatic detection of PPR becomes a highly unbalanced problem where PPR activity can be considered as anomalous activity surrounded by large portions of normal brain activity. This research proposes the use of a Variational Autoencoder (VAE) to extract features from the EEG segments by reducing them into the latent space to train several Anomaly Detection models proposed by the literature and compare their PPR detection performance. Results show that no AD model could detect any PPR activity correctly while the VAE model, also used for this task in our previous research, reached Accuracy, Sensitivity, and Specificity values of around 85%.
Martins et al. (Fri,) studied this question.