Authors :
Presenting Author: Ajay Thomas, MD, PhD – Baylor College of Medicine/Texas Children's Hospital
Sergio Morant Gálvez, MSc – NeuroBell
Robert R. Clancy, MD – Departments of Neurology and Pediatrics, University of Pennsylvania, and the Children's Hospital of Philadelphia, Philadelphia, PA
Mark O'Sullivan, PhD – NeuroBell
Alison O'Shea, PhD – NeuroBell
Rationale:
Neonatal seizures are an emergency, but most are undetectable by clinical observation alone, therefore electroencephalogram (EEG) monitoring is required to diagnose and quantify them. Neonatal seizures are typically focal in onset and can arise from any brain region. Their detection by EEG depends on the number and locations of scalp electrodes. Due to small neonatal head sizes, delicate skin, and the need for reduced handling, there are practical limitations on the total number of usable electrodes. This study investigates how the performance of a machine learning (ML) seizure-detection algorithm varies when, in the context of a fixed number of 9 recording electrodes, a standard bipolar montage is augmented to include extra, synthesized, “non-physiologic” bipolar channels.
Methods:
An ML model was trained on a proprietary expertly annotated dataset. The unseen test dataset was annotated by three experts who reviewed 18 standard bipolar channels derived from recording 19 electrodes. Tests in this study utilized 9 of these recording electrodes (Fig. 1) in the following configurations:
• Montage A: Ordinal pairs (8 channels)
• Montage B: Limited double banana (12 channels)
• Montage C: Expanded double banana (14 channels)
• Montage D: Full nearest neighbors (18 channels)
Montage A and B utilize conventional bipolar channels. Additional derivations were computed by re-referencing existing electrodes, maintaining the same number of electrodes. Many of these channels would be considered “non-physiologic” since they pair seemingly unrelated brain regions (Fig. 1D).
The Area Under the Curve (AUC) for each montage is reported in Fig. 2. The AUC is reported in comparison with a ground truth defined by both the majority voting and the consensus of three experts. The AUC is reported as an average value for the infants who experienced seizures (N=46 majority voting; N=39 consensus). The average Fleiss’ Kappa for each montage is reported in Fig. 2; this is an estimate of the interrater agreement across multiple experts and is reported on the entire cohort (N=79).
Results:
Model performance varied as the number of bipolar channels increased, despite the available referential EEG data remaining the same. For each montage pair, the p-value for AUC performance was calculated. A statistically significant improvement was found between montage A and C (p=0.044; consensus), indicating that this expanded montage improved seizure detection performance despite the fixed referential channels available.
Conclusions:
Electrographic seizures are recognized as discrete, pathophysiologic events in which ictal patterns first appear, then evolve in frequency, amplitude, morphology and usually location. These reflect the dynamic migrations of their underlying electrical fields during seizures. This study shows that by including even “non-physiologic” bipolar channels, seizure detection can be improved without additional electrodes. This approach to improving seizure detection has not been reported before and deserves further confirmation and study.
Funding: Southwest Pediatric Device Consortium (SWPDC) Seed Grant