Abstracts

Rationale:
There have been recent improvements in the treatments for epilepsy, ranging from anti-epileptic drugs and epilepsy surgery. However, these options may not only be costly and hard to access in parts of the world, but they have also been shown to be ineffective for some individuals with epilepsy. One major concern for people with untreated epilepsy is how to detect seizure episodes during their daily lives, as compared to the standard method of seizure detection in clinical settings via electroencephalographs (EEG). As such, this study aims to investigate whether more accessible electrocardiogram (ECG) data and Heart Rate Variability (HRV) metrics can be used to predict the onset of seizure events in patients with epilepsy.




Methods:
ECG and video EEG data was collected from 111 patients with focal-onset and generalized-onset epilepsy, totaling 3850 hours and 3281 interictal episodes. Data was gathered from the Tokyo Medical and Dental University Medical Hospital, the National Hospital Organization Nara Medical Center, National Center of Neurology and Psychiatry Hospital, Tohoku University Hospital, Osaka University Hospital Epilepsy Center, and University of Tokyo Hospital. The mean age of the patients was 31 ± 16 years, with 57 males and 54 females. Video, ECG, and EEG data of patients were simultaneously recorded for 24–72 hours using a long-term video-EEG monitoring system (Neurofax EEG-1200, NIHON KOHDEN). These tests were conducted in epilepsy monitoring units, and the sampling frequency of ECG and EEG was 500 or 1000 Hz. Using EEG data, Japan Epilepsy Society-certified epileptologists identified seizure events. Non-seizure data in the form of R-R Interval (RRI) data was used to train a self-attention autoencoder (SA-AE), a neural network that attempts to reproduce outputs that are similar to its inputs by calculating reconstruction error and an attention matrix. By calculating the average reconstruction error on non-seizure data and setting a reconstruction error threshold, the SA-AE was able to begin identifying anomalies in seizure data, which would have reconstruction errors higher than the set threshold. To correct the high false positive rate (FPR) associated with the SA-AE, anomalous data detected by the SA-AE was used to calculate 12 different HRV metrics, including time-domain, frequency-domain, and nonlinear metrics. These metrics were passed as the inputs through one of five types of binary classifiers: Random Forest, XGBoost, Logistic Regression, K-Nearest Neighbors, or Gradient Boosted Model.




Results:
While the SA-AE model alone had a sensitivity of 0.77 and a FPR of 0.66 times/hour, the combined SA-AE and logistic regression model had a sensitivity of 0.69 and a FPR of 0.30 times/hour. Additionally, of the data passed through both the SA-AE and binary classifier model, more than half of the patients had a FPR of 0 times/hour.




Conclusions:
Our results demonstrate the improvement of the SA-AE model with the binary classifier, maintaining a similar level of sensitivity while also decreasing the FPR by more than half. This may contribute to development of portable seizure detection technology outside of clinical settings via ECG data.




Funding:


Japan AMED Grant #21445838