Diagnosing Epileptogenesis with Deep Anomaly Detection

We propose a general framework for diagnosing brain disorders from Electroencephalography (EEG) recordings, in which a generative model is trained with EEG data from normal healthy brain states to subsequently detect any systematic deviations from these signals. We apply this framework to the early diagnosis of latent epileptogenesis prior to the first spontaneous seizure. We formulate the early diagnosis problem as an unsupervised anomaly detection task. We first train an adversarial autoencoder to learn a low-dimensional representation of normal EEG data with an imposed prior distribution. We then define an anomaly score based on the number of one-second data samples within one hour of recording whose reconstruction error and the distance of their latent representation to the origin of the imposed prior distribution exceed a certain threshold. Our results show that in a rodent epilepsy model, the average reconstruction error increases as a function of time after the induced brain injury until the occurrence of the first spontaneous seizure. This hints at a protracted epileptogenic process that gradually changes the features of the EEG signals over the course of several weeks. Overall, we demonstrate that unsupervised learning methods can be used to automatically detect systematic drifts in brain activity patterns occurring over long time periods. The approach may be adapted to the early diagnosis of other neurological or psychiatric disorders, opening the door for timely interventions.