Foundation Models
13 ECG foundation models and 3 PPG foundation models with code and weights available.
| Model | Year | Type | Backbone | Method | Pretrain Data | Data Size | Links |
|---|
Compare Model Compute
Computational benchmarks across different hardware. Click model names to view full details including architecture.
Hardware: A100-SXM4-80GB, Tesla T4 16GB, Intel Xeon @ 2.20GHz (4 cores, 50GB RAM) on Google Colab.
Benchmarks: Batch size 32 (GPU) / 8 (CPU), with adaptive batch sizing for OOM (32→16→8→4→2→1). Each model uses its native input specification.
Finetune hours estimated for 100K samples × 10 epochs.
* HeartLang uses 12 leads but is pre-tokenized to (B, 256, 96) via QRS-Tokenizer.
† GFLOPs: thop does NOT correctly count matrix multiplications inside nn.MultiheadAttention (specifically q @ k and attn @ v operations). Transformer models may be undercounted.
‡ S4 (supervised) is a baseline end-to-end supervised model from Al-Masud et al.
| Model | Leads | Params (M) | GFLOPs† | Infer (ms) | Throughput (/s) | Train (ms/sample) | Finetune (hrs) | Infer Mem (GB) | Train Mem (GB) |
|---|
12-Lead ECG Datasets
Publicly available 12-lead ECG datasets used in foundation model research.
| Dataset | Records | Patients | Country | Setting | Access | Pretrain | Eval | Link |
|---|
Reduced-Lead ECG & PPG Datasets
Datasets with single-lead ECG, reduced leads, or PPG signals.
| Dataset | Records | PPG | ECG | Country | Setting | Access | Pretrain | Eval | Link |
|---|
Compare Models
Select 2-4 models to compare side by side.