Verification of Heart Rate Variability and Autoencoder Based Seizure Prediction Algorithm Using Large Clinical Database

Abstract

Epileptic patients have a risk of accidents associated with seizures, which can lead to severe injury or death. If patients can predict seizures before a seizure onset, they can prevent such accidents by ensuring their safety. We have developed an algorithm to predict seizures by detecting changes in heart rate patterns observed before seizure onsets. In our previous research, heart rate variability (HRV), which is the fluctuations of an RR interval (RRI), is monitored using autoencoder (AE) for seizure prediction. However, our previous studies have limitations: a small amount of clinical data included in the study. In this study, we report the verification results of our developing seizure prediction algorithm using clinical data collected from 180 epileptic patients. As a result, the sensitivity of 77.9% and the area under the ROC curve (AUC) of 0.91 were achieved. Although seizure prediction was effective in most patients in our algorithm, a few patients experienced many false positives. We will investigate the characteristics of such patients based on their medical records.