Supervised Transfer Learning for Personalized Heart Rate Based Epileptic Seizure Detection

Abstract

Seizure alarm systems can improve the quality of life of refractory epilepsy patients, but require seizure detection algorithms. State-of-the-art heart rate-based algorithms often use a patient-independent approach due to insufficient annotated patient data, which does not allow a robust personalization. Ictal heart rate changes are however patient-dependent and could benefit from personalized algorithms. In this study, we propose to personalize seizure detection by using supervised transfer learning, which allows to train a classifier with a limited amount of data by using a reference classifier. It is evaluated on 207 hours of data including 74 seizures from 6 patients. An optimal performance of 89.8% sensitivity was achieved with on average 1.1 false alarms per hour, which is 54% false alarms less than the reference patient-independent classifier by using a limited amount of patient data. This shows that transfer learning can be used for a fast and robust personalization of detection algorithms.