Development of Multimodal Automated Seizure Detection and Type Classification Using Synchronized Video-EEG Data
Nan Lin1, Qiang Lu1, Weifang Gao1, Heyang Sun1, Yuan Gao1, Zi Liang2, Lian Li2, Peng Hu2
1Peking Union Medical College Hospital, 2NetEase Media Technology Co.
Objective:
To develop and validate a multimodal transformer network for automated seizure detection and seizure type classification on a large-scale multicenter datasets, using video-EEG data, addressing the critical need for efficient epilepsy diagnosis.
Background:
Visual seizure analysis by EEG specialists are time-consuming and labor-intensive. Deep learning based automated seizure recognition is significant for clinical practice. However, most researches with excellent seizure detection performance conducted on datasets less than 30 patients, offering limited coverage of seizure heterogeneity and between-patient variability. Seizure type classification studies face label scarcity, particularly for comprehensive focal seizure classification, and lack of external validation.
Design/Methods:
We included 820 patients, 5674 seizures from three epilepsy centers for model development and validation. Patient-wise splits and external center dataset were used for testing, to evaluate the robustness of model across individuals and centers. The model employed pretrained convolutional networks to extract features from EEG and video data. Transformer fuses these heterogeneous data modalities, enabling unified seizure analysis that addresses both detection and nine-class type classification (typical absence, atypical absence, epileptic spasm, myoclonic, tonic, temporal, frontal, centro-parietal and occipital seizures) tasks.
Results:
Validating on strict patient-wise split test set containing 151 patients, vEpiSzNet achieves robust seizure detection (AUROC 0.990, AUPRC 0.838). For seizure type classification, external 74 patients (583 seizures ) validation results showed 69.2% accuracy. Of the 74 patients, 44 patients were selected for clinician-machine comparison, resulting comparable performance on seizure classification between vEpiSzNet and EEG specialist (73.4% vs 73.7%). Machine showed dramatically reduced analysis time compared to clinician analysis (minutes versus hours). Video integration reduces false positives by 24.3% and improves frontal seizure classification by 38.4%. External validation across centers confirms generalizability (AUROC 0.923, AUPRC 0.607).
Conclusions:
We developed vEpiSzNet, a multimodal pretrained Transformer that trained and validated on a large-scale multicenter datasets, shows excellent performance in seizure detection and classification achieving specialist-level accuracy.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.