Model reference — AI Medical · Neurology

AI models for EEG interpretation and sleep staging

Sleep staging is the mature part of this field: published agreement with human scorers sits around 0.79 F1 and 85 percent, close to the ceiling set by inter-scorer disagreement, and the models are resilient to which channels you have.

EEG foundation models are newer and less settled. They produce embeddings across arbitrary montages, which is what makes them useful where your labelled set is small, and their reported accuracy is task-dependent by nature.

Every model on this page runs as part of a managed AI pipeline in our GPU clusters, with a dedicated private cluster in our cloud or an on-premise installation where medical governance requires it. Output is decision support for a qualified professional to review, not a diagnosis.

EEG interpretation and sleep staging service AI Medical services Pricing

Input type — EEG and polysomnography signals

Models in this group take multi-channel biosignal recordings; the rate is recorded hours processed per hour of wall-clock time. Each table gives three hardware tiers — Minimum, the smallest configuration on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is what one server of that tier processes per hour. Use these figures for initial sizing only. Before production we benchmark your own data to confirm accuracy, latency, throughput and cost.

U-Sleep

Vendor: University of Copenhagen

What it does: stages a whole night at 0.79 F1 and 0.76 kappa across fifteen public cohorts — the level two human scorers agree at — and it is resilient to which channels you have.

Mean F1 about 0.79, Cohen kappa about 0.76 across 15 public sleep cohorts; resilient to channel choice (independent, npj Digital Medicine).

Requirement	Minimum	Medium	High
GPU type	RTX 3060	RTX 4090	L40S 48 GB
VRAM	12 GB	24 GB	48 GB
vCPUs	4	8	16
RAM	16 GB	32 GB	64 GB
Server	1× RTX 3060 12 GB	1× RTX 4090 24 GB	1× L40S 48 GB
Rate (audio hours/hour)	≈ 200	≈ 700	≈ 1,800

YASA

Vendor: UC Berkeley (Walker lab)

What it does: stages sleep from a single channel and also marks spindles, slow waves and artefact, at about 85 percent agreement with human scorers.

About 85 percent agreement with human scorers on held-out cohorts (independent, eLife 2021).

Requirement	Minimum	Medium	High
GPU type	RTX 3060	RTX 4090	L40S 48 GB
VRAM	12 GB	24 GB	48 GB
vCPUs	4	8	16
RAM	16 GB	32 GB	64 GB
Server	1× RTX 3060 12 GB	1× RTX 4090 24 GB	1× L40S 48 GB
Rate (audio hours/hour)	≈ 200	≈ 700	≈ 1,800

LaBraM

Vendor: Tsinghua University

What it does: is the strongest EEG foundation model on abnormal-EEG detection and event classification, producing embeddings from any montage.

State of the art on abnormal-EEG detection, event classification and emotion tasks, balanced accuracy 0.66–0.82 by task (independent, ICLR 2024).

Requirement	Minimum	Medium	High
GPU type	RTX 3060	RTX 4090	A100 80 GB
VRAM	12 GB	24 GB	80 GB
vCPUs	6	12	24
RAM	16 GB	32 GB	64 GB
Server	1× RTX 3060 12 GB	1× RTX 4090 24 GB	1× A100 SXM 80 GB
Rate (audio hours/hour)	≈ 90	≈ 315	≈ 810

EEGPT

Vendor: EEGPT authors

What it does: produces EEG embeddings for decoding and classification tasks trained on your own labels.

Task-dependent across EEG decoding and classification benchmarks.

Requirement	Minimum	Medium	High
GPU type	RTX 3060	RTX 4090	A100 80 GB
VRAM	12 GB	24 GB	80 GB
vCPUs	6	12	24
RAM	16 GB	32 GB	64 GB
Server	1× RTX 3060 12 GB	1× RTX 4090 24 GB	1× A100 SXM 80 GB
Rate (audio hours/hour)	≈ 90	≈ 315	≈ 810

Neuro-GPT

Vendor: Neuro-GPT authors

What it does: is an EEG foundation model for downstream prediction where your labelled set is small.

Task-dependent across EEG downstream tasks.

Requirement	Minimum	Medium	High
GPU type	RTX 3060	RTX 4090	A100 80 GB
VRAM	12 GB	24 GB	80 GB
vCPUs	6	12	24
RAM	16 GB	32 GB	64 GB
Server	1× RTX 3060 12 GB	1× RTX 4090 24 GB	1× A100 SXM 80 GB
Rate (audio hours/hour)	≈ 90	≈ 315	≈ 810

BIOT

Vendor: BIOT authors

What it does: handles EEG, ECG and other biosignals in one model, which avoids a separate encoder per signal type in a mixed pipeline.

Task-dependent across EEG, ECG and biosignal classification tasks.

Requirement	Minimum	Medium	High
GPU type	RTX 3060	RTX 4090	A100 80 GB
VRAM	12 GB	24 GB	80 GB
vCPUs	6	12	24
RAM	16 GB	32 GB	64 GB
Server	1× RTX 3060 12 GB	1× RTX 4090 24 GB	1× A100 SXM 80 GB
Rate (audio hours/hour)	≈ 90	≈ 315	≈ 810

Choosing between them

For sleep, pick on channel availability and cohort: both leading models are cheap enough to run on every study. For everything else, the foundation models plus a small classifier of your own are the practical route, since no released model covers a clinical EEG label set end to end.

Accuracy figures above are those the producers and independent evaluations report, on their own test sets. They are a shortlist tool, not a prediction of what you will see. At the start of a project we run a short proof of concept on a sample of your own data, which replaces them with real figures — so the cost and the schedule for the full engagement are known before anything is committed.

EEG interpretation and sleep staging service AI Medical services Pricing

From benchmark to production

Send a representative sample, your expected volume and your latency target for eeg interpretation and sleep staging. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster — with the cost per unit of work known before you commit.