Rationale:
Long-term wearable monitoring is essential for reliable seizure tracking and treatment evaluation. While EEG is the gold standard for inpatient monitoring, its cumbersome, intrusive, and stigmatizing nature limits extended practical use. ECG offers a more discreet alternative, providing valuable seizure insights [1]. Despite frequent co-recording, ECG is less often analyzed independently. Deep learning (DL) [2] is a powerful tool for large dataset analysis and successful seizure detection [3]. This study quantifies and compares seizure-related information in ECG and EEG using DL.
Methods:
We developed a versatile DL model for offline seizure detection, processing either a single ECG or two EEG derivations. This model combines a Convolutional Neural Network and a Transformer Decoder to predict 10-second seizure probabilities. Input signals (2 min 10 sec, 256 Hz) include 1-minute pre/post context around the 10-second chunk. Two separate models were trained (an EEG model on 2 EEG channels; an ECG model on 1 ECG channel). A final combined EEG+ECG model was obtained by averaging their predictions.
Our study used the SeizIt2 dataset [4]: 2,853 records from 125 focal epilepsy patients (11,648 hours total, 886 annotated seizures/14 hours). Each record includes one ECG derivation and two EEG derivations (from T3-T4, T3-T5, T4-T6 options). A 5-fold cross-validation scheme with 25 patients/fold (3 folds for training, 1 fold for validation, 1 fold for final testing) was employed. Performance was evaluated using Precision-Recall (PR) curves and Area Under the PR curve (AUC-PR), confidence intervals are reported in brackets.
Results:
The EEG model outperformed the ECG model, with its PR curve dominating, almost everywhere (Fig 1). The EEG model achieved an AUC-PR of 0.0786 (0.0402 - 0.1129) while the ECG model achieved an AUC-PR of 0.0405 (0.0270 - 0.0796). Importantly, the combined EEG+ECG approach surpassed single-modality performance with an AUC-PR of 0.0854 (0.0645-0.1560), indicating complementary information.
Conclusions:
While ECG carries seizure-related information usable by DL, its standalone detection performance is insufficient when compared to 2-channel EEG models. Nevertheless, their observed complementarity suggests significant promise for robust combined detection. This combined approach inspired the solution that ranked 1st in the UNA Europa Seizure Detection Challenge (Feb-May 2025).
References:
[1] Verrier RL et. al. Epileptic heart: A clinical syndromic approach. Epilepsia. 2021;62:1780–1789.
[2] LeCun Y et al. Deep learning. Nature. 2015;521:436–444.
[3] Tveit J et al. Automated Interpretation of Clinical Electroencephalograms Using Artificial Intelligence. JAMA Neurol. 2023;80(8):805–812. doi:10.1001/jamaneurol.2023.1645
[4] Bhagubai M et al. SeizeIT2. OpenNeuro [Dataset]. 2025.
Funding:
This work was funded by Reliev Technologies.