Authors :
Presenting Author: Mao Otake, Student – Tokyo Medical and Dental University, National Center of Neurology and Psychiatry, Japan
Kentaro Hori, Technician – Quadlytics. Inc; Rikumo Ode, Graduate – Department of Material Engineering, Nagoya University; Koichi Fujiwara, Associate professor – Department of Material Engineering, Nagoya University; Motoki Inaji, Associate professor/Lecturer – Department of Neurosurgery, Tokyo Medical and Dental University; Taketoshi Maehara, Professor – Department of Neurosurgery, Tokyo Medical and Dental University; Toshitaka Yamakawa, Professor – NARA Institute of Science and Technology; Manabu Kano, Professor – Department of Systems Science, Graduate School of Informatics, Kyoto University; Miho MIyajima, Associate professor/Lecturer – Department of Psychiatry and Behavioral sciences, Tokyo Medical and Dental University
Rationale:
We developed machine learning algorithms to predict epileptic seizures [1] based on heart rate variability analysis [2]. Previously, we trained and validated the model using electrocardiogram (ECG) signals during long-term video electroencephalography (vEEG) monitoring. This study assessed the feasibility of using a wearable ECG monitoring system for seizure prediction.
Methods:
This study included four patients with temporal lobe epilepsy who underwent long-term vEEG or stereotactic electroencephalography (SEEG) monitoring. These patients had a median age of 32.5 years, and 50% were women. Concurrent ECG recordings with a wearable ECG device were performed during long-term vEEG or SEEG monitoring. Figure 1 shows the employed wearable device, equipped with two recording electrodes positioned anatomically between the ventricular septum and the left ventricular anterior wall. In each patient, we simultaneously collected and analyzed ECG waveforms recorded with the EEG as a component of long-term EEG monitoring (E-ECG) and ECG waveforms recorded with the wearable device (W-ECG). We evaluated the robustness of artifacts in each ECG recording by calculating the robustness index as ([Total recording time of ECG] − [Total recording time of artifacts])/[Total recording time of ECG]. Artifacts were detected using denoising autoencoder algorithms. Thereafter, the performance of the seizure prediction algorithms on E-ECG and W-ECG data was compared. The algorithms based on a neural network for time-series data [3] applied to the E-ECG and W-ECG data for each patient. Seizures were predicted if a preictal anomaly of HRV was detected within 15 minutes before clinical onset. Moreover, we calculated the sensitivity, false positive (FP) rates, FP numbers per hour, and the area under the curve (AUC) of the receiver operating characteristic curve.
Results:
The median recording duration of E-ECG was 67 h and 30 min, with a range of 44 h and 28 min to 97 h and 1 min. The robustness indexes were 67.8% for E-ECG and 75.8% for W-ECG on average (Fig. 2). After excluding artifacts, the E-ECG and W-ECG recordings contained twelve and eleven preictal periods, respectively. Seizure types included focal to bilateral tonic-clonic, focal impaired awareness, and subclinical seizures. Regarding the performance of the algorithms, W-ECG achieved a sensitivity of 63.6%, marginally surpassing 61.5% for E-ECG. The overall false positive frequency per hour for W-ECG was 0.69, which was reduced from 1.34 observed for E-ECG. The total AUC score for W-ECG was 0.710, whereas for E-ECG, an AUC score of 0.623 was observed.Conclusions:
The robustness of artifacts in W-ECG is comparable with that in E-ECG. Moreover, the performance of our seizure prediction algorithms for W-ECG data was equal to or better than that for E-ECG. This study suggests the feasibility of the wearable epileptic seizure prediction system based on our HRV-based algorithms. Further studies with a larger cohort in outpatient settings are crucial to determine the clinical potential and usefulness of the system.
Funding:
This work was supported in part by Japan Agency for Medical Research and Development (AMED) Grant Number 21445838.