Skip to main content

Model reference — AI Medical · Cardiology

AI models for ECG interpretation

Two kinds of model are in play. Foundation models produce embeddings and are the better choice when your labelled set is small — a linear probe on top reaches useful accuracy with a few hundred examples. Diagnostic models return probabilities across a fixed label set directly.

Accuracy in this area is genuinely task-dependent, and most producers publish no single figure. Where one is reported, treat it as a shortlisting signal, not a specification: ECG model performance drops predictably between the population it was trained on and yours.

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.

ECG interpretation service AI Medical services Pricing

Input type — ECG waveform

Models in this group take an ECG waveform, typically ten seconds at 500 Hz. 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.

ECGFounder

Vendor: Harvard Medical School / Emory University, PKU Digital Health and collaborators

What it does: scores 150 diagnostic labels from one waveform, in both single-lead and 12-lead form — the broadest label coverage on this page, at AUROC around 0.95 internally and 0.90 externally.

Mean AUROC about 0.95 across 150 diagnostic labels on internal test, about 0.90 external (developer-reported).

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 80 GB
VRAM12 GB24 GB80 GB
vCPUs61224
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (waveforms/hour)≈ 18,000≈ 63,000≈ 162,000

HeartBEiT

Vendor: Icahn School of Medicine at Mount Sinai

What it does: reads the ECG as an image and beats conventional networks when labelled data is scarce, which is the usual situation for a specific local endpoint.

AUC 0.90–0.93 for LV dysfunction and hypertrophic cardiomyopathy at low label budgets; beats CNNs when training data is scarce (independent, npj Digital Medicine 2023).

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 80 GB
VRAM12 GB24 GB80 GB
vCPUs61224
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (waveforms/hour)≈ 18,000≈ 63,000≈ 162,000

ECG-FM

Vendor: University Health Network / Vector Institute, Bo Wang Lab

What it does: produces waveform embeddings that reach AUROC 0.83–0.94 across ejection fraction, arrhythmia and outcome tasks with only a linear probe on top — a few hundred labels rather than a few hundred thousand.

AUROC 0.83–0.94 across LVEF, arrhythmia and outcome tasks after linear probing (independent).

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 80 GB
VRAM12 GB24 GB80 GB
vCPUs61224
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (waveforms/hour)≈ 18,000≈ 63,000≈ 162,000

ECGFM-KED

Vendor: ECGFM-KED authors

What it does: brings external diagnostic knowledge into the embedding, for classification where the label set is defined clinically rather than statistically.

Task-dependent; evaluated on multiple ECG downstream tasks, no single universal accuracy.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 80 GB
VRAM12 GB24 GB80 GB
vCPUs61224
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (waveforms/hour)≈ 18,000≈ 63,000≈ 162,000

HuBERT-ECG

Vendor: HuBERT-ECG authors

What it does: adapts self-supervised speech modelling to ECG, producing embeddings for downstream cardiac tasks.

Task-dependent; reported across multiple cardiac applications, no single universal accuracy.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 80 GB
VRAM12 GB24 GB80 GB
vCPUs61224
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (waveforms/hour)≈ 18,000≈ 63,000≈ 162,000

ST-MEM

Vendor: ST-MEM authors

What it does: is a representation model for ECG, intended as the base layer under your own classifier.

Task-dependent; representation-learning results vary by downstream ECG dataset.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 80 GB
VRAM12 GB24 GB80 GB
vCPUs61224
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (waveforms/hour)≈ 18,000≈ 63,000≈ 162,000

CLEF-Large

Vendor: Nokia Bell Labs

What it does: is the largest of three single-lead encoders from the same family — the one to use when wearable and patch data is your input and accuracy governs.

Task-dependent; evaluated on multiple single-lead ECG diagnostic benchmarks.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 80 GB
VRAM12 GB24 GB80 GB
vCPUs61224
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (waveforms/hour)≈ 18,000≈ 63,000≈ 162,000

CLEF-Medium

Vendor: Nokia Bell Labs

What it does: trades a little accuracy for a smaller footprint, for single-lead work at volume.

Task-dependent; evaluated on multiple single-lead ECG diagnostic benchmarks.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090L40S 48 GB
VRAM12 GB24 GB48 GB
vCPUs4816
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× L40S 48 GB
Rate (waveforms/hour)≈ 40,000≈ 140,000≈ 360,000

CLEF-Small

Vendor: Nokia Bell Labs

What it does: is the lightest single-lead encoder in the family, cheap enough to run on continuous wearable streams.

Task-dependent; evaluated on multiple single-lead ECG diagnostic benchmarks.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090L40S 48 GB
VRAM12 GB24 GB48 GB
vCPUs4816
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× L40S 48 GB
Rate (waveforms/hour)≈ 40,000≈ 140,000≈ 360,000

MOMENT-1-base

Vendor: Auton Lab / Carnegie Mellon University

What it does: is a general time-series model rather than a cardiac one, which makes it useful when ECG is only part of a wider signal pipeline including anomaly detection and forecasting.

Task-dependent; general time-series foundation model evaluated on ECG classification among other tasks.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090L40S 48 GB
VRAM12 GB24 GB48 GB
vCPUs4816
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× L40S 48 GB
Rate (waveforms/hour)≈ 40,000≈ 140,000≈ 360,000

Choosing between them

The lead count decides most of it: 12-lead hospital ECG and single-lead wearable data need models built for them. Beyond that, if you hold labels, a diagnostic model is quicker to deploy; if you do not, an embedding model plus a small classifier of your own is the cheaper path to a working endpoint.

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.

ECG interpretation service AI Medical services Pricing

From benchmark to production

Send a representative sample, your expected volume and your latency target for ecg interpretation. 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.