Authors :
Presenting Author: Qiufan Chen, BS – Kyoto University
Koichi Fujiwara, Dr. – Nagoya University; Miho Miyajima, MD – Tokyo Medical and Dental University; Kentaro Hori, MS – Quadlytics Inc.; Motoki Inaji, MD – Tokyo Medical and Dental University; Taketoshi Maehara, MD – Tokyo Medical and Dental University; Masaki Iwasaki, MD – National Center of Neurology and Psychiatry; Ayataka Fujimoto, MD – Seirei Christopher University; Satoshi Maesawa, MD – Nagoya University; Manabu Kano, PhD – Kyoto University
Rationale:
For individuals with refractory epilepsy, uncontrolled seizures can be life-threatening. When conventional drugs fail, this study seeks to predict seizures using RR interval (RRI) on an electrocardiogram (ECG), as seizures are known to impact heart rate patterns before their onset. Compared to the previous study (R. Ode et al. Artificial Life and Robotics, 28(2), 403-409, 2023), we have optimized more hyperparameters, and evaluated a new group of epileptic patients.
Methods:
A self-attentive autoencoder (SA-AE) model is constructed using clinical RRI data. SA-AE is a neural network that incorporates the self-attention (SA) mechanism within the hidden layer of the autoencoder (AE). SA captures dependencies and relationships among various RRI sequence elements. Additionally, multi-head attention is used in SA to increase training stability. AE is a type of neural network model commonly used to minimize the reconstructed error (RE) between the reconstructed input and the original input. Hence, AE trained by only interictal RRIs will not be able to reconstruct preictal RRIs. By monitoring the RE, we can detect heart rate pattern changes before the seizure onset.
Results:
Clinical data were collected from 122 focal epilepsy patients, totaling 1828 hours, approved by the Ethics Committee of Tokyo Medical and Dental University and five other organizations. Of these patients, 58 were female, and 64 were male, with an average age of 30.7 ± 16.0. Rather than splitting the dataset based on episodes as before, this study opted to split the dataset based on individual patients. Training data only includes the interictal RRI data, while validation and testing datasets include both interictal and preictal data. The hyperparameters of the model were optimized by grid search. In this study, successful seizure prediction means forecasting a seizure within 20 minutes of its onset. Fig. 1 illustrates an example of seizure prediction achieved by monitoring the RE. The pink band represents anomalies detected in heart rate patterns and the blue band represents the cooldown period of the system, indicating successful prediction. The performance on the testing data demonstrated a sensitivity of 82%, a false positive rate of 0.68 times per hour, and an area under the receiver operating characteristic curve of 0.94.
Conclusions:
In this study, the seizure prediction system showed promising results on previously unseen patient data, offering insights into its future reliability for new patients. In future works, we will continue to optimize the system and work on bringing the epilepsy prediction system into operation.
Funding:
Japan Agency for Medical Research and Development (AMED) Grant Number 21445838