MORGOTH
Toward unified and comprehensive automated electroencephalogram interpretation: a multicentre development and validation of an electroencephalogram foundation model
Clinical Question
Can a single EEG foundation model deliver expert-level interpretation across the full range of clinical EEG tasks and settings?
Bottom Line
MORGOTH, a unified EEG foundation model, delivers expert-level performance (AUC-ROC 0.86–0.98) across 17 clinically relevant EEG findings and all major clinical settings, with robust external validation and reduced age/sex sensitivity compared with prior task-specific models — supporting deployment for improved diagnostic access in low-resource settings and greater efficiency in high-volume centres.
Major Points
- First EEG foundation model providing expert-level interpretation across routine outpatient, ICU, EMU, and sleep laboratory settings within a single architecture
- Achieved AUC-ROC 0.86–0.98 across 17 EEG findings including seizures, epileptiform discharges, IIIC patterns, pathological slowing, and sleep staging
- Outperformed ≥90% of experts on 3 of 7 multi-expert-annotated datasets and exceeded ≥20% of experts on each task
- Event-level EUC=96.6% for seizure/IIIC detection and EUC=100% for spike detection
- External validation from 48 institutions showed only modest performance declines
- More robust to age and sex variation than SPaRCNet
Design
Study Type: Multicentre development and validation study of an AI foundation model
Randomization:
Enrollment Period: Jan 1, 2003 to Feb 1, 2025
Centers: 52
Countries: USA, Canada, Switzerland, Belgium, Brazil, China, UK, Greece
Sample Size: 33584
Analyzed: 33584
Analysis: Model performance evaluated using AUC-ROC, precision-recall curves, percentage of experts' operating points under the curve (EUC), inter-rater reliability (IRR), and statistical calibration; model–expert vs expert–expert agreement compared
Inclusion Criteria
- Patients with clinical EEG recordings across major clinical settings
- Ages 0 to >90 years
- Recordings from routine outpatient clinics, epilepsy monitoring units, critical care, and sleep laboratories
Exclusion Criteria
- Non-clinical applications such as brain–computer interfaces or emotion recognition (excluded in comparator literature review, not necessarily patient exclusions)
Arms
| Field | MORGOTH foundation model | Control |
|---|---|---|
| N | 33584 | 0 |
| Intervention | Multidomain omnibus foundation model for EEG interpretation supporting event-level detection and EEG-level interpretation across routine, ICU, EMU, and sleep settings | Comparator: 6–30 experts per test dataset annotations; prior task-specific models including SCORE-AI, SPaRCNet, SpikeNet, U-Sleep |
| Duration |
Outcomes
| Outcome | Type | Control | Intervention | HR / OR / RR | P-value |
|---|---|---|---|---|---|
| Model performance across 17 clinically relevant EEG findings compared with expert consensus using AUC-ROC and EUC (percentage of experts' operating points under the curve) | Primary | Expert-level performance (multi-expert consensus) | AUC-ROC 0.86–0.98 across 17 EEG findings; outperformed ≥90% of experts on 3/7 multi-expert-annotated datasets; exceeded ≥20% of experts on each task | Event-level EUC=96.6% for seizure/IIIC detection; EUC=100% for spike detection | |
| External validation performance vs internal test sets | Secondary | Event-level AUC change –1.21%, EUC –3.33%; EEG-level AUC –2.12%, EUC –9.52% | |||
| Robustness across age | Secondary | MORGOTH age sensitivity 20.90% vs SPaRCNet 30.90% | |||
| Robustness across sex | Secondary | MORGOTH sex-related differences 33.33% vs SPaRCNet 44.00% | |||
| Inter-rater reliability analysis | Secondary | MORGOTH matched or exceeded expert consensus | |||
| Robustness to moderate channel loss | Secondary | Performance remained robust | |||
Subgroup Analysis
Performance evaluated across age, sex, and moderate channel loss; MORGOTH showed less age sensitivity and fewer sex-related differences than SPaRCNet
Funding
US National Institutes of Health
Based on: MORGOTH (Lancet Digital Health, 2026)
Authors: Sun C, Karakis I, Herlopian A, ..., Jing J
Citation: Sun C, et al. Lancet Digit Health. 2026. doi:10.1016/j.landig.2026.101039
Content summarized and formatted by NeuroTrials.ai.