MORGOTH

A hospital that wants machine help reading its EEGs has had to assemble it from parts: one model for the routine outpatient report, another for the ictal–interictal–injury continuum in the ICU, a third for interictal spikes, a fourth for overnight sleep staging. Each carries its own preprocessing assumptions and its own blind spots, and none covers the settings the others were built for.

MORGOTH — the authors' backronym for "multidomain omnibus for reading and generalising over thorough EEG interpretation" — collapses that toolkit into one model. It was developed at Beth Israel Deaconess Medical Center with Massachusetts General Hospital, Brigham and Women's Hospital and Boston Children's Hospital, and published in The Lancet Digital Health. A self-supervised tokenizer and transformer learn a general representation of scalp EEG; lightweight task heads then read that representation at two time scales at once, emitting per-segment event detections and whole-recording findings from the same frozen backbone.

LaBraM and EEGFormer pretrain on public and brain–computer-interface corpora and are scored on machine-learning benchmarks; MORGOTH trains on hospital recordings and is scored against practising electroencephalographers, with panels of six to 30 annotators per test set setting the inter-rater ceiling it is measured against.

#Key Features

• Seventeen findings from one backbone: One pretrained model supplies normal/abnormal, focal and generalised slowing, burst suppression, spike detection and localisation, seizures and the four ictal–interictal–injury continuum patterns, benign variants, and sleep stages.
• Two time scales in one pass: Fully connected heads handle event-level detection on 1-second and 10-second segments; convolutional and attention heads pool the same representation into whole-recording classifications.
• Raw clinical files as input: Unaltered EDF or MAT recordings go straight in; the model handles filtering, resampling, montaging, clipping, normalisation and epoching itself, imputing missing channels from the mean of those available.
• Trained on expert disagreement: Where several annotators labelled a sample, the vote distribution is used as a soft label, so the heads learn calibrated probabilities instead of collapsing genuine ambiguity onto one class.
• Evaluated against annotator panels: Performance is reported as experts-under-the-curve — the fraction of individual experts' operating points falling below the model's ROC or precision-recall curve — alongside AUC and Cohen's κ.

#Technical Details

The tokenizer converts continuous EEG into a codebook of 8,192 discrete tokens using contrastive learning and vector quantisation. On top of it sits a transformer with 12 encoder blocks, multi-head attention, and both temporal and spatial position encoding, so electrode geometry is part of the representation, not an afterthought. Input is 19-channel EEG in the international 10–20 system at 200 Hz, bandpass filtered 0.5–70 Hz, notch filtered at 50/60 Hz, clipped at ±500 μV and normalised on a common average montage; variable-length recordings are supported at inference. Pretraining is self-supervised over recordings from 14,500 patients in the Harvard Electroencephalography Database, after which the task heads are fine-tuned with focal loss and curriculum learning under AdamW.

Development drew on EEGs from 18,677 patients recorded between January 2003 and February 2025 across the four Boston hospitals. Internal validation covered 13,334 held-out patients; external validation covered 1,573 patients from 48 institutions that contributed no training data. AUC-ROC ran from 0.86 to 0.98 across the 17 findings, with experts-under-the-curve of 96.6% for seizure and ictal–interictal–injury continuum detection and 100% for spike detection. Moving from internal to external cohorts cost 1.21% AUC and 3.33% EUC at the event level, and 2.12% AUC and 9.52% EUC at the recording level. Age sensitivity (20.90% vs 30.90%) and sex-related differences (33.33% vs 44.00%) were lower than for the earlier SPaRCNet detector.

#Applications

The clearest use is triage and second-read support where fellowship-trained electroencephalographers are scarce: flagging seizures and periodic patterns during continuous ICU monitoring, pre-screening routine outpatient studies, marking interictal spikes for epilepsy workup, and staging polysomnography in adult and paediatric sleep laboratories — the last benchmarked against sleep-specific systems such as U-Sleep. Because one model covers all of these, a department runs one inference pipeline rather than four, with outputs as calibrated probabilities thresholded to local workload. The released code also serves as a feature extractor, with downstream work using its event-level activations to derive EEG severity measures.

#Impact

By spanning the clinical EEG settings in one model and scoring itself against annotator panels rather than a single reference standard, MORGOTH sets a useful bar for how automated EEG interpretation gets evaluated: beating one rater's labels on one corpus says little about a reading room where the experts themselves disagree. The limits are real. Validation is retrospective, with no prospective or deployment trial reported; the code carries a CC BY-NC 4.0 licence prohibiting commercial use; and the pretrained checkpoints and the labelled corpus sit behind the Brain Data Science Platform's credentialed-access process rather than being openly downloadable. Channel-loss robustness is qualified to moderate loss, so the model is not montage-agnostic.