Automated Neonatal Seizure Detection: Improved Accuracy And Reduced False Alarms.pdf

This study presents an improved automated system for detecting neonatal seizures from continuous EEG monitoring in NICUs. The challenge lies in the non-stationary nature of neonatal EEG signals and the need to distinguish seizures from artifacts.

Key Points:

1. Existing Challenges:
- Visual recognition of neonatal seizures is labor-intensive, inconsistent among raters, and requires specialized expertise.
- Current automated systems have high false alarm rates.

2. Proposed Solution:
- An enhanced heuristic seizure detection model combining spike-train and oscillatory detectors.
- A post-processing technique using support vector machine (SVM) and mean phase coherence (MPC) to reduce false alarms.

3. Methodology:
- The heuristic model uses Teager-Kaiser nonlinear energy (TKE) and discrete wavelet transform (DWT) for feature extraction.
- The SVM classifier is trained on a dataset of 17 neonates, each with extensive EEG-polygraphy recordings.
- MPC is calculated between EEG and polygraphic signals to characterize artifacts.

4. Results:
- False alarm rate drops by 42% (from 2.6 to 1.5 false detections per hour).
- Good detection rate decreases only slightly by 4%.

5. Significance:
- The multi-stage approach combining heuristic and data-driven models offers a transparent, interpretable solution for automated neonatal seizure detection.
- This advancement supports clinical decision-making in NICUs by alleviating workload and improving care for vulnerable newborns.