An article in Nature highlights a breakthrough in scaling convolutional neural networks (CNNs) to detect seizures in neonatal electroencephalograms (EEGs) with expert-level precision.
Study: Scaling convolutional neural networks achieves expert level seizure detection in neonatal EEG. Image Credit: Gorodenkoff/Shutterstock.com
The researchers demonstrated how advancements in CNN architecture and training can enable accurate, automated seizure detection, a critical tool for early intervention in newborns. Their findings showcase the potential of scaled CNNs as reliable aids in neonatal care, reducing dependence on manual EEG analysis.
Background
Neonatal seizures are a pressing concern in the first few days of life, often caused by hypoxic-ischaemic encephalopathy (HIE) or cerebrovascular injuries. These seizures, which typically occur within the first 72 hours, are difficult to detect due to their subtle presentation and the scarcity of continuous EEG monitoring expertise. Early detection is essential, as untreated seizures can result in severe outcomes, including death or long-term neurological deficits.
While automated seizure detection tools have shown promise, they remain limited by small datasets, reliance on global rather than per-channel annotations, and inadequate validation on independent datasets. The researchers addressed these limitations by developing a deep-learning-based CNN specifically designed for per-channel seizure detection. Their approach leverages larger datasets, modern architectures, and rigorous validation, aiming to bridge gaps in neonatal EEG analysis.
Developing and Validating the Model
The study focused on creating an AI-powered model for seizure detection in term neonates using EEG data. The researchers analyzed recordings from 202 neonates at Cork University Maternity Hospital (CUMH) in Ireland, covering both at-risk and healthy infants. This dataset included 50,299 hours of EEG, during which 12,402 seizure events were identified and validated by neurophysiologists with high inter-rater agreement.
To account for variability in EEG setups and seizure patterns, the researchers adopted the ConvNeXt CNN architecture, adapted for one-dimensional (1D) time-series data. The model was trained on 16-second EEG segments, using techniques like dynamic undersampling and loss re-weighting to address the inherent imbalance between seizure and non-seizure data. Robustness was further enhanced through data augmentation strategies, including flip and cutout transformations.
The model’s performance was validated on two independent datasets: one from CUMH and another open-access dataset from Helsinki. Metrics such as false detections per hour and seizure burden per hour were used to evaluate performance. To ensure clinical relevance, the researchers also tested the model’s equivalence to human experts, comparing its predictions to inter-rater agreement among neurophysiologists.
Pre- and post-processing steps, such as bandpass filtering, downsampling, and artifact removal, ensured high-quality data inputs. These measures emphasized the model’s potential to transform neonatal seizure detection and improve outcomes in neonatal intensive care units (NICUs).
Key Findings and Insights
The researchers systematically scaled both model size and training data to evaluate their impact on seizure detection. CNNs ranged from a 39,000-parameter "nano" variant to a 21-million-parameter "extra-large" (XL) model, with larger models consistently outperforming smaller ones. Performance metrics, including Matthews correlation coefficient (MCC) and error rate, showed significant improvement, with gains validated across datasets from Cork and Helsinki.
Increasing the training dataset size also led to significant performance gains, with up to a 50 % improvement observed when datasets were 20 times larger than those used in prior research. The XL model achieved expert-level accuracy, effectively identifying seizures while maintaining a low false detection rate for neonates without seizures. It also reliably estimated seizure burden, a critical metric for clinical decision-making.
However, performance declined slightly on the Helsinki dataset for the largest models, attributed to differences in clinical protocols and neonatal age. When the data was split into early- and late-EEG groups, performance scaled predictably for early EEGs but degraded for late EEGs, suggesting age-specific features in EEG data. Despite this, the XL model demonstrated resilience to channel loss, with minimal performance drops even when 50 % of EEG channels were removed.
Significance and Limitations
This study highlighted the value of scaling both training data and model complexity for neonatal seizure detection. The ConvNeXt-based XL model, with its 21 million parameters, set a new benchmark in neonatal EEG analysis, outperforming smaller models while generalizing effectively across datasets.
The researchers underscored the limitations of common metrics like the area under the receiver-operating-characteristic curve (AUC) in handling imbalanced datasets. Instead, they prioritized balanced metrics such as MCC and Cohen’s κ, which better captured performance in seizure detection tasks. The use of per-channel annotations further enhanced adaptability to various clinical settings, offering a more granular approach than previous global annotation methods.
Despite its strong performance, the model faced challenges. For example, diminishing returns were observed with the largest models, particularly for late EEG data. The study’s reliance on single-center data and reduced generalization for neonates beyond the first week of life also highlighted areas for future improvement.
Conclusion and Future Directions
This research demonstrates the potential of scaling CNNs for neonatal seizure detection, achieving expert-level accuracy through advanced AI techniques. By leveraging large datasets and sophisticated model architectures, the team developed a scalable and robust tool for reliable seizure monitoring across diverse clinical environments.
While limitations such as single-center data and performance variability with late EEGs remain, the findings underscore AI's promise in improving neonatal care. Future research will aim to validate these models using multi-center datasets and refine their performance across broader clinical contexts. With continued development, this technology could enable real-time, automated seizure monitoring, enhancing outcomes for neonates in intensive care settings.
Journal Reference
Hogan et al., 2025. Scaling convolutional neural networks achieves expert level seizure detection in neonatal EEG. Npj Digital Medicine, 8(1). DOI:10.1038/s41746-024-01416-x https://www.nature.com/articles/s41746-024-01416-x
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.