P2. Unsupervised anomaly detection for diagnosing brain disorders from EEG recordings – Results from a rodent epilepsy model

We present a general machine learning (ML) framework for diagnosing brain disorders using EEG recordings. It could detect any systematic deviations from the baseline (BL) EEG recordings. We test the approach using a rodent model of an acquired mesial temporal lobe epilepsy (mTLE) via electrical perforant pathway stimulation (PPS) [1] followed by usually a latent period of several weeks of epileptogenesis (EPG) progression without any seizures. Identifying ongoing EPG during this latent period prior to the first spontaneous seizure might enable early intervention. Previous supervised approaches have proven that deep neural networks can distinguish intracranial EEG signals between the BL period prior to PPS and the EPG period after PPS [2]. Here, we formulate the problem as an unsupervised anomaly detection task, which allows learning the statistics of the input data without the need for any labels as in supervised approaches. Specifically, we first train an autoencoder (AE) to learn to encode EEG data from the BL period. We then apply the trained AE on data from the EPG period. We expect that activity patterns systematically deviating from those observed during BL should be characterized by higher reconstruction errors (REs). Indeed, our results show that REs averaged over one hour gradually increase as a function of time after the PPS (Fig. 1). We also find that the distribution of REs from the BL and the EPG phases are significantly different (Fig. 1B). EEG segments with particularly high REs exhibit spikes and high frequency oscillations (Fig. 1B, right inset). Interestingly, there are an increased large number of segments with unusually low REs (Fig. 1B, left inset), which exhibits EEG slowing and delta rhythm (1.5–3 Hz). In conclusion, we showed that unsupervised ML with AEs allows the automatic detection of systematic drifts in brain activity, which may facilitate rapid EEG recordings for early diagnosis of brain disorders and open the door for early interventions. References Costard, L.S., Neubert, V., Veno, M.T., Su, J., Kjems, J., Connolly, N.M., Prehn, J.H., Schratt, G., Henshall, D.C., Rosenow, F. and Bauer, S., 2019. Electrical stimulation of the ventral hippocampal commissure delays experimental epilepsy and is associated with altered microRNA expression. Brain Stimulation, 12(6), pp. 1390-1401. Lu, D., Bauer, S., Neubert, V., Costard, L.S., Rosenow, F. and Triesch, J., 2020, September. Towards Early Diagnosis of Epilepsy from EEG Data. In Machine Learning for Healthcare Conference (pp. 80-96). PMLR. Fig. 1. A. The reconstruction error evolution throughout the whole recording period of rat #1B. The distribution of the reconstruction errors from the BL and the EPG phases from the recordings of rat #1. We also show some typical EPG segments whose reconstruction errors are below the 10th percentile and over the 90th percentile, highlighted with a solid and a dashed black box, respectively. Other stimulated rats show a similar trend.