Personalizing Heart Rate-Based Seizure Detection Using Supervised SVM Transfer Learning.

Area of Science:

• Biomedical Engineering
• Neurology
• Signal Processing

Background:

• Automated seizure detection is crucial for wearable seizure warning systems to enhance the quality of life for refractory epilepsy patients.
• Current patient-independent algorithms for heart rate-based seizure detection suffer from patient-specific ictal heart rate variations, leading to high false detection rates.
• Limited availability of annotated patient data hinders the development of fully patient-specific seizure detection algorithms.

Purpose of the Study:

• To introduce a novel transfer learning approach for personalizing heart rate-based seizure detection using minimal patient data.
• To evaluate the efficacy of this personalized approach in reducing false detections compared to patient-independent methods.
• To assess the adaptability and robustness of the transfer learning method for patient-specific seizure detection.

Main Methods:

• A transfer learning approach was developed to personalize heart rate-based seizure detection algorithms.
• The algorithm was trained and evaluated using 2,172 hours of single-lead ECG data from 24 temporal lobe epilepsy patients.
• The dataset included 227 focal impaired awareness seizures, with a focus on using only a few days of data per patient for personalization.

Main Results:

• The personalized approach achieved an overall sensitivity of 71% with 1.9 false detections per hour.
• A 37% average decrease in the false detection rate was observed compared to the reference patient-independent algorithm.
• The transfer learning method demonstrated faster and more robust adaptation to patient-specific characteristics than existing personalization alternatives.

Conclusions:

• The proposed transfer learning method offers an easily implementable solution for personalizing heart rate-based seizure detection.
• This personalization can significantly improve the quality of life for refractory epilepsy patients when integrated into multimodal seizure detection systems.
• The approach effectively addresses the challenge of limited patient data for developing accurate, personalized seizure detection algorithms.