ECG/PPG Open Foundation Models and Datasets

A comprehensive collection of open-source foundation models and publicly available datasets for electrocardiogram (ECG) and photoplethysmogram (PPG) analysis.

🌐 Interactive Version: https://aramis00.github.io/ECGPPG_FM_dataset/

Overview

This repository provides curated lists of:

• 17 Foundation Models for ECG/PPG analysis (12-lead ECG, single-lead ECG, and PPG)
• 15 Open 12-Lead ECG Datasets
• 23 Open Reduced-Lead ECG and PPG Datasets
• Computational Benchmarks comparing model inference/training performance across hardware

All models include links to code repositories and pretrained weights. All datasets include access links and citations.

Data Files

For programmatic access, data is available in CSV and JSON formats:

data/
├── models.csv                    # Foundation models
├── models.json
├── ecg_datasets_12lead.csv       # 12-lead ECG datasets
├── ecg_datasets_12lead.json
├── ecg_ppg_datasets_reduced.csv  # Reduced-lead ECG & PPG datasets
└── ecg_ppg_datasets_reduced.json

Quick Start (Python)

import pandas as pd

# Load models
models = pd.read_csv('data/models.csv')

# Load datasets
datasets_12lead = pd.read_csv('data/ecg_datasets_12lead.csv')
datasets_reduced = pd.read_csv('data/ecg_ppg_datasets_reduced.csv')

# Filter by size (using numeric columns)
large_datasets = datasets_12lead[datasets_12lead['records_numeric'] > 100000]

---

Foundation Models

12-Lead ECG Foundation Models

Model	Year	Backbone	Method	Pretrain Data	Data Size	Code	Weights
ECG-JEPA	2024	Transformer	M/R/G	CSN, Code-15	180K	Code	Weights
HuBERT-ECG	2024	CNN+Transformer	M/R/G	CODE, CPSC, CPSC-Extra, PTB, PTB-XL, ...	9.1M	Code	Weights
DeepECG	2025	CNN+Transformer	CL	MIMIC-IV-ECG, Code-15, MHI-ds*	1.9M	Code	Weights
ECG-FM	2025	CNN+Transformer	CL	CPSC, CPSC-Extra, PTB-XL, Georgia, CS...	870K	Code	Weights
HeartLang	2025	Transformer	M/R/G	MIMIC-IV-ECG	800K	Code	Weights
ECG-CPC	2025	CNN+SSM	CL	HEEDB	10.7M	Code	Weights
ESI	2024	CNN (Text: Transformer)	CL+M/R/G	PTB-XL, CSN, MIMIC-IV-ECG	660K	Code	Weights
MERL	2024	CNN (Text: Transformer)	CL	MIMIC-IV-ECG	771K	Code	Weights
MELP	2025	CNN+Transformer (Text: Transformer)	CL+M/R/G	MIMIC-IV-ECG	760K	Code	Weights
KED	2024	CNN (Text: Transformer)	CL	MIMIC-IV-ECG	800K	Code	Weights
ECGFounder	2024	CNN	SL (multilabel)	HEEDB	10.7M	Code	Weights
ST-MEM	2024	Transformer	M/R/G	CSN, Code-15	189K	Code	Weights

Single-Lead ECG Foundation Models

Model	Year	Backbone	Method	Pretrain Data	Data Size	Code	Weights
ECG-PT	2024	Transformer	M/R/G	PhysioNet CinC 2020	42M tokens	Code	Weights
HeartBERT	2024	Transformer	M/R/G	MIT-BIH, PTB-XL, European ST-T	72M tokens	Code	Weights

PPG Foundation Models

Model	Year	Backbone	Method	Pretrain Data	Data Size	Code	Weights
PPG-PT	2024	Transformer	M/R/G	Capnobase/BIDMC/Cuffless BP	128M tokens	Code	Weights
PaPaGei-S/-P	2024	CNN	CL	VitalDB, MIMIC-III waveform, MESA sleep	57K hours	Code	Weights
PulsePPG	2025	CNN	CL+M/R/G	MOODS	55K hours	Code	Weights

---

Open Datasets

Access Legend

• O = Open (freely available)
• R = Restricted (requires Data Use Agreement)
• C = Credentialed (requires credentialing process)

12-Lead ECG Datasets

Dataset	Records	Patients	Country	Setting	Access	Link
HEEDB	11.7 M	2.1 M	US	hospital diagnostic ECG	C	Link
CODE	2.3 M (346 K CODE-15%)	1.7 M(234 K CODE-15%)	Brazil	tele-health setting	R (CODE 15%, O; CODE-test, O)	Link
MIMIC-IV ECG	800 K	161 K	US	hospital diagnostic ECG	C	Link
SNUH LYDUS 167K	167 K/, 50K	167 K/ 50K	South Korea	Hospital diagnostic ECG	R (50K, O)	Link
IKEM	98.1 K	30.3 K	Czech Republic	hospital diagnostic ECG	O	Link
UK Biobank	91.1 K	82.7 K	UK	volunteered exams, diagnost...	R	Link
CSN	45.2 K	45.2 K	China	hospital diagnostic ECG	O	Link
SPH	25.8 K	25.7 K	China	hospital diagnostic ECG	O	Link
PTB-XL	21.8 K	18.9 K	Germany	diagnostic ECG - NR-Schille...	O	Link
ZZU pECG	14.2 K	11.6 K	China	hospital diagnostic ECG, pe...	O	Link
Georgia*	10.3 K	10.2 K	US	hospital diagnostic ECG	O	Link
CPSC 2018, CPSC 2018 extra	10.3 K	9.5 K	China	hospital diagnostic ECG	O	Link
SaMi-Trop	2.0 K	1.6 K	Brazil	Chagas cohort	O	Link
Pre-/Post-STEMI ECG Database, University of Michigan	266	266	US	hospital diagnostic ECG, ST...	O	Link
EchoNext	100K	36.3K	US	hospital diagnostic ECG	O	Link

Reduced-Lead ECG and PPG Datasets

Dataset	Records	PPG	ECG	Country	Access	Link
MIMIC-III Waveform Database	67.8K (MIMIC-III match...	✓	✓	US	O	Link
WAVES	550K (ECG); 510K (PPG)	✓	✓	US	R	Link
SCOPE	4.5K	✓	✓	South Korea	O	Link
VitalDB	6.4K	✓	✓	South Korea	O	Link
MOVER	83.5K	✓	✓	US	O (currently N/A due to data revision)	Link
MIT-BIH Arrhythmia	48	-	✓	US	O	Link
Long-term ST	86	-	✓	Slovenia, Italy	O	Link
European ST-T	90	-	✓	8 Europe countries	O	Link
St Petersburg INCART	75	-	✓	Russia	O	Link
MIT-BIH Atrial fibrillation	25	-	✓	US	O	Link
Long term AF	84	-	✓	US	O	Link
IRIDIA-AF	167	-	✓	Belgium	O	Link
CPSC 2021 Paroxysmal Afib	1.4K	-	✓		O	Link
Icentia11k	54K	-	✓	Canada	O	Link
CapnoBase	42	✓	✓	Canada	R	Link
MESA polysomnography	2K	✓	✓	US	R	Link
SDB	146	✓	✓	Canada	O	Link
NuMoM2b	5337	✓	✓	US	R	Link
PPG-BP	657	✓	-	China	O	Link
PPG DaLiA	15	✓	✓	Germany	O	Link
WESAD	15	✓	✓	Germany	O	Link
TROIKA (= IEEEPPG)	12	✓	✓	China	O	Link
ECSMP	89	✓	✓	China	O	Link

---

Computational Benchmarks

We provide comprehensive computational benchmarks comparing foundation models on parameters, FLOPs, inference speed, training throughput, and memory usage across different hardware (A100, T4, CPU).

script/FM_computation/
├── FM_computation.py              # Benchmark script
├── FM_computation.md              # Documentation & results
└── results/
    ├── ecg_benchmark_A100_*.json  # A100 GPU results
    ├── ecg_benchmark_T4_*.json    # T4 GPU results
    └── ecg_benchmark_cpu_*.json   # CPU results

Key findings:

• Training time is dominated by sequence length and FLOPs, not parameter count
• HuBERT-ECG (92.8M params) trains ~3× faster than ECG-FM (90.4M params) due to more aggressive downsampling (93 vs 312 transformer tokens)
• MERL (ResNet18) achieves fastest inference at 9,616 samples/sec on A100
• Training memory scales with sequence × hidden dim: ESI (ConvNeXtV2-Base) is the largest training-memory consumer in the suite (~16 GB on A100)

See script/FM_computation/FM_computation.md for detailed benchmark results and methodology.

---

Abbreviations

Pretraining Methods

• CL = Contrastive Learning
• M/R/G = Masked/Reconstructive/Generative Learning
• SL = Supervised Learning
• SSL = Self-Supervised Learning

Technical Terms

• ECG = Electrocardiogram
• PPG = Photoplethysmogram
• AUROC = Area Under the Receiver Operating Characteristic Curve
• Dx. = Diagnosis
• Afib = Atrial Fibrillation
• LVEF = Left Ventricular Ejection Fraction

Datasets

• HEEDB = Harvard-Emory ECG Database
• CODE = Clinical Outcomes in Digital Electrocardiography
• MIMIC = Medical Information Mart for Intensive Care
• PTB-XL = Physikalisch-Technische Bundesanstalt ECG Database
• CSN = Chapman-Shaoxing-Ningbo Database
• CPSC = China Physiological Signal Challenge

---

Citation

If you use this resource, please cite the relevant papers for each model and dataset you use. BibTeX citations are available in the JSON data files.

---

Contributing

Contributions are welcome! Please open an issue or pull request to:

• Add new open-source foundation models
• Add new publicly available datasets
• Fix errors or update links

---

License

This repository contains metadata and links only. Please refer to individual model and dataset licenses for usage terms.