Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Using Explainable AI for EEG-based Reduced Montage Neonatal Seizure Detection

Dinuka Sandun Udayantha, Kavindu Weerasinghe, Nima Wickramasinghe, Akila Abeyratne, Kithmin Wickremasinghe, Jithangi Wanigasinghe, Anjula De Silva, Chamira U. S. Edussooriya

TL;DR

This work tackles neonatal seizure detection under practical constraints by proposing an explainable deep learning model that operates on a reduced $12$-channel EEG montage. The architecture blends a CNN encoder for temporal features, a Graph Attention Network for spatial channel interactions, and a small MLP classifier, with post-hoc Grad-CAM-style interpretability to pinpoint time windows and channels driving decisions. On the Helsinki Zenodo dataset, the method achieves notable gains over state-of-the-art reduced-montage baselines in AUC and recall during 10-fold cross-validation, and demonstrates real-time inference performance suitable for NICU deployment. The approach promises accessible, interpretable seizure monitoring in resource-limited settings and points to future enhancements via artifact removal and self-supervised learning to broaden generalization.

Abstract

The neonatal period is the most vulnerable time for the development of seizures. Seizures in the immature brain lead to detrimental consequences, therefore require early diagnosis. The gold-standard for neonatal seizure detection currently relies on continuous video-EEG monitoring; which involves recording multi-channel electroencephalogram (EEG) alongside real-time video monitoring within a neonatal intensive care unit (NICU). However, video-EEG monitoring technology requires clinical expertise and is often limited to technologically advanced and resourceful settings. Cost-effective new techniques could help the medical fraternity make an accurate diagnosis and advocate treatment without delay. In this work, a novel explainable deep learning model to automate the neonatal seizure detection process with a reduced EEG montage is proposed, which employs convolutional nets, graph attention layers, and fully connected layers. Beyond its ability to detect seizures in real-time with a reduced montage, this model offers the unique advantage of real-time interpretability. By evaluating the performance on the Zenodo dataset with 10-fold cross-validation, the presented model achieves an absolute improvement of 8.31% and 42.86% in area under curve (AUC) and recall, respectively.

Using Explainable AI for EEG-based Reduced Montage Neonatal Seizure Detection

TL;DR

This work tackles neonatal seizure detection under practical constraints by proposing an explainable deep learning model that operates on a reduced -channel EEG montage. The architecture blends a CNN encoder for temporal features, a Graph Attention Network for spatial channel interactions, and a small MLP classifier, with post-hoc Grad-CAM-style interpretability to pinpoint time windows and channels driving decisions. On the Helsinki Zenodo dataset, the method achieves notable gains over state-of-the-art reduced-montage baselines in AUC and recall during 10-fold cross-validation, and demonstrates real-time inference performance suitable for NICU deployment. The approach promises accessible, interpretable seizure monitoring in resource-limited settings and points to future enhancements via artifact removal and self-supervised learning to broaden generalization.

Abstract

The neonatal period is the most vulnerable time for the development of seizures. Seizures in the immature brain lead to detrimental consequences, therefore require early diagnosis. The gold-standard for neonatal seizure detection currently relies on continuous video-EEG monitoring; which involves recording multi-channel electroencephalogram (EEG) alongside real-time video monitoring within a neonatal intensive care unit (NICU). However, video-EEG monitoring technology requires clinical expertise and is often limited to technologically advanced and resourceful settings. Cost-effective new techniques could help the medical fraternity make an accurate diagnosis and advocate treatment without delay. In this work, a novel explainable deep learning model to automate the neonatal seizure detection process with a reduced EEG montage is proposed, which employs convolutional nets, graph attention layers, and fully connected layers. Beyond its ability to detect seizures in real-time with a reduced montage, this model offers the unique advantage of real-time interpretability. By evaluating the performance on the Zenodo dataset with 10-fold cross-validation, the presented model achieves an absolute improvement of 8.31% and 42.86% in area under curve (AUC) and recall, respectively.
Paper Structure (11 sections, 3 equations, 4 figures, 1 table)

This paper contains 11 sections, 3 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Proposed Model Architecture
  3. Deep Learning Model Architecture
  4. CNN Encoder
  5. GAT Network
  6. Classification Head
  7. Model Interpretability
  8. Dataset and Pre-processing
  9. Model Training
  10. Results and Discussion
  11. Conclusion

Figures (4)

  • Figure 1: The proposed deep learning model architecture. The first $4$ blocks belong to the CNN-based temporal feature extractor. The $12$ channels are preserved throughout the CNN encoder but down-sampled the temporal features at the end of each block using Average Pooling. GAT layers $1, 2,$ and $3$ have the output shapes $(12\times 37)$, $(12\times 32)$, and $(12\times 16)$ respectively and the last multilayer perceptron (MLP) network has $32, 16$, and $1$ neurons, respectively.
  • Figure 2: (a) The proposed reduced montage electrode placement for seizure detection on the international $10-20$ system (b) The illustration of the employed reduced montage graph representation of the selected electrode montage. The graph nodes represent the channels and the edges represent the functional connectivity between channels.
  • Figure 3: Performance comparison between GAT layers and scaled dot product attention layers. Training for dot product attention was terminated after $50$ epochs due to low performance.
  • Figure 4: Top subplot: Comparison between true labels and model prediction probabilities. Next $12$ subplots: Visualization of $7.5$ minutes EEG, where the recording is observed to be seizure free up to $4$ mins $25$ secs, and a seizure occurs past this point. Last subplot: A zoomed-in version for better visualization of seizure onset and how the relevance changes.