Multimodal Epileptic Seizure Detection System – SeizeIT2
Published:
Duration: 2024 – 2025
Role: Algorithm developer for multimodal fusion and model compression.
Collaborators: International collaboration with KU Leuven, UCL, Aarhus University, and clinical sites across Europe (EU multi-center clinical project).
Overview
The SeizeIT2 project is an international multicenter clinical trial (NCT04284072) — the first-ever phase-4 clinical trial for a wearable device targeting in-home seizure monitoring. The project unites public and private stakeholders across Europe to address an unmet clinical need: reliable, continuous seizure detection using non-invasive wearable sensors in real-world settings.
This project is a continuation of the SeizeIT1 ICON project, expanding the scope to a much larger patient cohort and multimodal sensor setup.
Developed attention-based multimodal deep learning algorithms for neurophysiological signal analysis, with cross-modal feature fusion strategies to improve seizure detection accuracy and robustness. Explored model compression techniques to enable deployment on resource-constrained edge devices.
Clinical Partners
Data was collected across 5 European Epilepsy Monitoring Units (EMUs):
| Center | Country |
|---|---|
| University Hospital Leuven (incl. pediatric patients) | Belgium |
| Freiburg University Medical Center | Germany |
| RWTH University of Aachen | Germany |
| Karolinska University Hospital | Sweden |
| Coimbra University Hospital | Portugal |
Data collection spanned from January 2020 to June 2022, approved by the UZ Leuven ethics committee (approval ID: S63631).
Dataset
| Property | Detail |
|---|---|
| Patients | 125 (51 female, 41%) with focal epilepsy |
| Total recording time | ~11,640 hours of wearable data |
| Total seizures | 886 focal seizures |
| Mean seizure duration | 58 seconds (range: 3 sec – 16 min) |
| Modalities | 4 (behind-the-ear EEG, ECG, EMG, accelerometer/gyroscope) |
| Environment | Hospital video-EEG monitoring |
Multimodal Wearable Sensor Setup
The recording setup consists of two Sensor Dot (SD) wearable devices:
- Behind-the-ear EEG (bte-EEG): Two Ag/AgCl cup electrodes placed on the mastoid bone (corresponding to T7/T8 and P9/P10 positions), creating two bipolar channels — one per hemisphere.
- ECG: Two electrodes on the left chest measuring cardiac activity.
- EMG: Two electrodes on the left deltoid muscle capturing muscle activity.
- Motion (ACC + GYRO): Both SD modules contain built-in accelerometers and gyroscopes for movement monitoring.
All participants have wearable bte-EEG recordings. In approximately 3% of cases, ECG/EMG/motion data were unavailable due to technical issues. Concurrent video-EEG monitoring with a standardized 25-electrode array served as the gold-standard reference for seizure annotation.
Seizure Distribution
By Seizure Type
| Type | Count | Percentage |
|---|---|---|
| Focal Aware (FA) | 317 | 35.8% |
| Focal Impaired Awareness (FIA) | 393 | 44.4% |
| Focal-to-Bilateral Tonic-Clonic (FBTC) | 55 | 6.2% |
| Focal with unclear awareness | 12 | 1.4% |
| Subclinical focal | 2 | 0.2% |
| Unknown / unreported onset | 93 | 10.5% |
By Hemisphere Onset
| Hemisphere | Percentage |
|---|---|
| Left | 44% |
| Right | 12% |
| Bilateral | 1% |
| Unclear | 43% |
Seizure onsets were distributed across the central, frontal, temporal, occipital, parietal, and insular lobes, with a predominance of temporal lobe seizures (30%). About 26% of seizures could not be paired with a clear onset lobe.
Deliverables
- Participated in the SeizeIT2 Seizure Detection Challenge.
- Lightweight seizure detection model optimized for edge deployment.
- Presented results at multiple consortium meetings.
- Attended and presented at the final SeizeIT2 workshop.