AI Is Going to Eat Basic Healthcare. This EEG Model Is the Receipt.

This article is educational and is not medical advice. It does not diagnose, treat, prevent, or cure any condition. Komori is not a medical device, is not FDA-cleared, does not detect seizures, and is not intended for people with epilepsy. Talk with a neurologist or epileptologist about personal medical decisions. Households living with epilepsy should use clinically validated seizure-detection devices as directed by their care team.

It's 2 a.m. in the ICU. A patient's EEG has been running for six hours. Someone needs to look at it, interpret the squiggles, decide if the patterns are normal background, benign variants, or something that changes the treatment plan.

That someone is usually asleep at home, because there are not enough fellowship-trained EEG readers to staff every ICU bed in real time. The bottleneck isn't the data. It's the people who know how to read it.

AI was supposed to fix this years ago. It hasn't, mostly because the models we've had so far are narrowly scoped — one detects spikes, another handles sleep staging, a third is trained only on outpatient clinic recordings and falls apart in the ICU. Nobody built a model that could handle all of it at once, and nobody validated the thing against actual expert consensus rather than just "better than the last model."

On June 2, MORGOTH 1.0 went live on the Brain Data Science Platform.

MORGOTH stands for Multi-domain Omnibus for Reading and Generalizing Over THorough EEG interpretation. Terrible acronym, honestly. But it's a clinical EEG foundation model trained on 18,677 patients across four hospitals, validated internally on 13,334 patients and externally on 1,573 patients from 48 institutions, with performance measured not just against a test set but against panels of up to 30 human experts [^1].

And it works across the board: routine outpatient EEGs, epilepsy monitoring units, ICU continuous monitoring, and sleep labs. Event-level detection (seizures, spikes, rhythmic patterns) and whole-recording interpretation. Ages 0 to 90+. AUC-ROC scores of 0.86 to 0.98 across 17 EEG findings [^1].

The paper is in press at Lancet Digital Health. The data and code are on GitHub under CC BY-NC 4.0. This is not a press release. This is the receipt.

The scarce thing is expert time

Most EEGs are interpreted by physicians without specialized fellowship training, because there aren't enough fellowship-trained readers to go around. EEG misinterpretation is the leading cause of epilepsy misdiagnosis [^1].

That's not a technology problem. That's a people problem. And it's not going to be solved by training more people — the pipeline for fellowship-trained neurophysiologists is measured in years, and the demand is measured in millions of recordings per year.

So the question becomes: can a machine do the routine interpretation well enough that the human expert's time is spent only on the genuinely hard calls?

MORGOTH's answer appears to be yes, for a lot of the routine work.

What it actually does

MORGOTH handles 7 event-level tasks and 17 whole-recording classifications.

Event-level means it's scanning segments of EEG — typically 10-second or 1-second windows — and detecting specific patterns:

• Normal vs abnormal
• Burst suppression
• Spike detection (1-second resolution)
• Slowing (focal vs generalized vs none)
• Spike localization (focal vs generalized vs none)
• Sleep staging (5-class AASM-defined: wake, N1, N2, N3, REM)
• Seizure and ictal-interictal continuum patterns (seizure, LPD, GPD, LRDA, GRDA, other)

Whole-recording means it looks at the full EEG and outputs binary classifications for 17 different findings — things like "this recording contains epileptiform discharges" or "this shows pathological slowing."

The model was developed on data from four hospitals, then validated on independent datasets from 48 institutions. External validation showed modest performance drops — event-level AUC dropped 1.2%, whole-recording AUC dropped 2.1% — but held up [^1].

Here's the part that matters: on multi-expert datasets, MORGOTH outperformed an average of 90% of experts on 5 of 7 tasks, and exceeded at least 20% of experts on all 17 tasks [^1].

Read that carefully. It's not "better than the average expert." It's "better than most experts, on most tasks."

For seizure detection, the model hit 95% experts-under-curve (EUC) — meaning 95% of individual expert operating points fell below the model's ROC curve. For spike detection, 100% EUC [^1].

Sleep staging is the bridge

One of the 7 event-level tasks MORGOTH handles is 5-class AASM sleep staging: wake, N1, N2, N3, and REM.

This matters because sleep staging is where clinical EEG meets consumer wellness. A sleep lab polysomnogram is a medical test. A contactless radar sleep monitor logging your position and movement through the night is a wellness device. But the underlying biology — the architecture of sleep, the transitions between stages, the patterns that distinguish restorative sleep from fragmented garbage — that's the same physiology.

MORGOTH isn't designed to run on a radar sensor in your bedroom. It runs on 19-channel clinical EEG at 200 Hz, with electrodes glued to your scalp. That is not a consumer product.

But it proves the concept: a foundation model trained on diverse clinical data can generalize across settings and tasks, including sleep. And the fact that it handles sleep staging alongside seizure detection and ICU monitoring means the same architecture could be adapted for different modalities — including the contactless, camera-free, non-wearable monitoring that Komori is being built to do.

Komori tracks sleep position, movement, bed-exit events, and room conditions (CO2, temperature, humidity, light, noise) using radar. It is not EEG. It does not measure brain activity. It does not diagnose anything. It is a general wellness product, and it will not tell you which sleep stage you were in at 3 a.m.

But it will tell you which position you held, how often you moved, and whether the room was working for you or against you. And that's the kind of routine monitoring that used to require a person in the loop — either a caregiver checking on you, or a wearable you had to remember to charge and strap on every night.

The parallel to MORGOTH is this: wellness monitoring doesn't need to be medical-grade to be useful. It needs to be consistent, accessible, and interpretable. The bottleneck isn't the data. It's making the data something you can act on without needing an expert to translate it for you every morning.

The limits are real

MORGOTH is not a product you can download and run on your laptop. The model is available on GitHub, but the data it was trained on is credentialed-access only, hosted on BDSP's S3 bucket ¹. You need to complete a Data Use Agreement to pull the datasets.

The paper is in press, not yet printed. The validation is strong, but this is version 1.0 of a research tool, not an FDA-cleared diagnostic device.

And critically, this is still a model that reads EEG. It does not replace the neurologist who decides what to do with the finding. It does not run the differential diagnosis when the EEG is ambiguous and the clinical picture is messy. It does not sit with the family and explain what the results mean.

That's the hard call. That's where the human still matters.

Where I want the human, where I don't

Here's my take, as someone building a contactless wellness monitor and watching this space closely.

I want a human neurologist in the loop for:

• Diagnosing epilepsy
• Deciding when to start or stop medication
• Interpreting an EEG in the context of a complex clinical picture
• Explaining risk to a patient or family
• Making the call when the data is ambiguous and the stakes are high

I do not need a human in the loop for:

• Logging my sleep position every night
• Flagging when I got out of bed three times instead of my usual once
• Telling me my room was 1,180 ppm CO2 at 4 a.m. and that's why I woke up groggy
• Running a preliminary screen on an ICU EEG to surface the recordings that need urgent expert review

The first list is diagnosis and treatment. The second list is monitoring and triage. And the second list is where AI is going to eat basic healthcare, because the economics make no sense otherwise.

There are not enough fellowship-trained neurophysiologists to read every EEG. There are not enough caregivers to check on every aging parent every hour through the night. There are not enough sleep specialists to interpret every night of sleep data for every person who wants to know if their sleep is actually restorative.

So the routine stuff — the monitoring, the logging, the "is this normal or does a human need to look at it" triage — that becomes automated. And the human expertise gets reserved for the genuinely hard calls.

MORGOTH is one example of that shift. It's not the only one, and it's not a product yet. But it's a foundation model with expert-level performance across clinical settings, with the data and code published, and that's a big deal.

Where this is going

Foundation models work because they're trained on diverse data and can generalize to new tasks. MORGOTH was trained on EEGs from routine clinics, ICUs, epilepsy monitoring units, and sleep labs. It handles event detection and whole-recording classification. It generalizes to external institutions ¹.

That's the model for what comes next in wellness tech: not a device that does one thing in one setting, but a system trained on diverse physiological signals that can generalize across contexts.

Komori is not MORGOTH. Komori is a contactless radar monitor for sleep position and nighttime wellness, not a clinical EEG interpreter. But the principle is the same: build something that works across the variation people actually have — different ages, different conditions, different rooms, different beds — and make it interpretable without needing an expert in the loop every morning.

The bottleneck has always been people. The data was always there. Now the models are catching up.

And the future of basic healthcare is: machines handle the routine, humans handle the hard calls, and nobody waits six hours for a fellowship-trained neurophysiologist to wake up and look at an EEG that an AI could have triaged at 2 a.m.

That's not a vision. That's the economics.

---

Sources

Footnotes

1. Sun, C., Karakis, I., Herlopian, A., et al. (2026). MORGOTH 1.0: A Foundation Model for Clinical EEG - Data and Code (version 1.0.0). Brain Data Science Platform. DOI: 10.60508/v7ky-5g40. Companion paper: Sun C, Jing J, et al. Toward Unified and Comprehensive Automated EEG Interpretation: Multi-center Development and Validation of an EEG Foundation Model. Lancet Digital Health, in press (2026). Data and code: github.com/bdsp-core/morgoth (CC BY-NC 4.0). Note: as of publication, some users have reported the GitHub repository returns a 404 error; the BDSP page lists this URL as the official code repository. ↩ ↩²