Incorporating spike correlations into an SVM-based neonatal seizure detector

In this paper, we have adapted a spike correlation (SC) method of neonatal EEG seizure detection, so that it can be directly incorporated into an SVM-based algorithm. To this end, we estimate several features based on the analysis of the smoothed non-linear energy operator (SNLEO). SNLEO features alone resulted in a median AUC of 0.963 (IQR 0.919-0.985). This AUC was significantly higher than with the original SVM-based method (p=0.024). The SNLEO method was significantly improved by incorporating a selected number of features from the SVM-based detector (p=0.002). Median AUC with this feature set was 0.981 (IQR 0.942-0.994). This study confirms, that incorporating SNLEO features adapted from the SC method significantly improve the performance of an SVM-based neonatal EEG seizure detector.