Validating an SVM-based neonatal seizure detection algorithm for generalizability, non-inferiority and clinical
Karoliina T Tapani1, Päivi Nevalainen2, Sampsa Vanhatalo3
1BABA Center, Children's Hospital and Pediatric Research Center, Helsinki University Hospital and University of Helsinki, Helsinki, Finland; Department of Neuroscience and Biomedical Engineering, Aalto University, Espoo, Finland; Brain Modelling Group, QIMR Berghofer Medical Research Institute, Australia; Department of Medical Physics, Kymenlaakso Central Hospital, Kotka, Finland.
Computers in Biology and Medicine
|April 5, 2022
Summary
Neonatal seizure detection algorithms (SDA) are nearing human expert performance. Validation shows consistent algorithm performance, with retraining achieving non-inferiority for clinical efficacy in seizure detection.
Area of Science:
- Neonatal neurology
- Medical device technology
- Computational neuroscience
Background:
- Neonatal seizures require accurate detection for timely intervention.
- Current seizure detection algorithms (SDA) approach human expert annotation benchmarks.
- Assessing generalizability, non-inferiority, and clinical efficacy is crucial for SDA deployment.
Purpose of the Study:
- To validate a neonatal seizure detection algorithm (SDA) on an independent dataset.
- To evaluate the SDA's generalizability, non-inferiority to human experts, and clinical efficacy.
- To determine if SDA performance is comparable to human expert annotation in neonatal EEG analysis.
Main Methods:
- Validation on an independent dataset of 28 neonates.
- Testing generalizability by comparing training and validation set performance (cross-validation).
- Assessing non-inferiority via inter-observer agreement between SDA and human experts.
- Evaluating clinical efficacy by comparing seizure burden quantification and identification of clinically significant periods.
Main Results:
- Algorithm performance remained consistent between training and validation sets (AUC, p > 0.05).
- Initial SDA output was inferior to human experts; retraining with diverse data achieved non-inferior performance (Δκ = 0.077).
- SDA demonstrated high accuracy in assessing seizure burden (89-93%) and identifying clinically relevant periods (87%).
Conclusions:
- The validated neonatal SDA demonstrates consistent performance and approaches human equivalence.
- Retraining enhanced the algorithm's non-inferiority, indicating improved generalizability and reliability.
- The SDA provides clinically relevant EEG interpretations, nearing human expert capabilities for neonatal seizure detection.
