Neural Noise

Optimal graph representations and neural networks for seizure detection using intracranial EEG

2 min read

  • Graph representations combined with neural networks enable automated seizure detection from intracranial EEG (iEEG) data.1
  • The methods are reusable for other neural time-series applications beyond seizure detection.1 1

Gardner updates

  • Optimal graph representations and neural networks for multichannel time series data improve seizure phase classification (Nature); methods transfer to BCI-relevant neural data analysis. 2

Weekly enrichment (2026-07-20)

  • The study evaluated four graph neural network (GNN) architectures for seizure detection from intracranial EEG (iEEG), with the top-performing model reaching a seizure detection accuracy of 97%.34
  • Two multi-center international iEEG datasets were used, comprising 23 and 16 patients respectively, with 5 and 7 hours of iEEG recordings.34
  • In the graph representation, nodes correspond to iEEG electrodes and edges encode functional connectivity metrics such as Pearson correlation, phase-locking value, and coherence.35
  • Adding average energy-per-electrode and frequency-band information as node features, together with frequency-specific coherence as edge features, substantially improved detection performance.35
  • The best binary classification configuration reported mean accuracy 97.07% (SD 2.62%), mean AUC 0.9942 (SD 0.0065), and mean F1-score 0.9616 (SD 0.0335).4
  • The highest-performing architecture combined a 2-layer design with an edge-conditioned convolution (ECC) layer followed by a graph attention network (GAT) layer.35
  • An earlier version of the pipeline (bioRxiv) used a publicly available OpenNeuro dataset of iEEG seizure recordings from 25 patients collected across 4 US epilepsy centers, reaching ~91% accuracy, 95% AUC, and 90% F1 for binary seizure detection.6
  • For multi-class classification (preictal vs. ictal vs. postictal), the optimized graph representations achieved mean accuracy ~90%, AUC ~97%, and F1 ~88% across subjects.46
  • The authors emphasize explainability (including t-SNE visualization of embeddings) as key to clinical trust, noting that despite >95% accuracy prior ML solutions have not been fully deployed clinically.43

Footnotes

  1. https://news.google.com/rss/articles/CBMie0FVX3lxTFAycVM5cGlHU1FoN0NoeWNmYWtMREp3ZkN0YlF1Qno4YVFmTDZ4TlpIazY5aGY4WjE1bXZaMlJJZFlFeC1sTEMyRXhfVV9OOUQ0RUlLRTFGZ18ySFAtOGc1YkdiSDU1TUhqWFhFcmxBY0xMU25tOXpzMjJ5QQ?oc=5 ↩ ↩2 ↩3

  2. https://news.google.com/rss/articles/CBMiX0FVX3lxTE5GUXU5RmxJa2h5eGlINnR4cUozREVmRnZzekFNR3VQcWVtTGM4UWhheDhDY0F5VFNmVW1KSUdLcmg1V29lMW82a0RaUGNnS1BmRXRMV3ZXSHh2OWgtRy0w?oc=5 ↩

  3. https://doi.org/10.1101/2024.12.28.24316703 ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  4. https://www.nature.com/articles/s41598-025-01882-7 ↩ ↩2 ↩3 ↩4 ↩5

  5. https://www.medrxiv.org/content/10.1101/2024.12.28.24316703v1.full-text ↩ ↩2 ↩3

  6. https://www.biorxiv.org/content/10.1101/2023.06.02.543277v1 ↩ ↩2

Graph View