HRV-based Monitoring of Neonatal Seizures with Machine Learning

Abstract

With the rapid development of machine learning (ML) in biomedical signal processing, ML-based neonatal seizure detection using heart rate variability (HRV) parameters extracted from the electrocardiogram (ECG) has gained increasing interest. In this paper, we present a benchmarking of various ML classifiers for HRV-based neonatal seizure monitoring. We extract the HRV parameter in time-domain, frequency-domain, and nonlinear-domain from segments with duration ranging from 30 to 180 s and perform the feature selection with minimum redundancy and maximum relevance (mRmR). In the next step, we evaluate the performance using nested cross-validation on a dataset collected from 16 preterm and term newborns with neonatal seizures with a total duration of over 35 hours. The best-performing classifier was the support vector machine (SVM) with a linear kernel using HRV parameters from the 180 s segment, achieving an area under the operator characteristic operating curve (AUC) score of 0.627, 89.7% sensitivity, 34.6% specificity, and 92.3% good detection rate.