ICML 2026 · Official PyTorch implementation
LVCG is the first general self-supervised framework for 12-lead ECG that learns in latent vectorcardiogram (VCG) space rather than raw lead signals. Because standard ECG is multiple linear views of the same cardiac field, lead-space learning entangles pathology with electrode geometry and generalizes poorly under domain shift; inspired by the Frank VCG model, LVCG instead recovers a unified 3D field and learns view-invariant representations via self-supervised multi-lead reconstruction—lifting visible leads, encoding beat morphology through a token bottleneck, modeling inter-beat dynamics, and projecting back to missing leads with a non-learnable geometry layer. This physically grounded design yields compact embeddings that transfer strongly to linear probing, multi-lead view reconstruction, and non-cardiac detection, especially with scarce labels and cross-site shift.
conda create -n lvcg python=3.10
conda activate lvcg
pip install -e .Requirements: PyTorch 2.x, NumPy, SciPy, PyYAML, WFDB (MIMIC loading), scikit-learn (probing), pandas (AI-READI).
PhysioNet datasets require credentialed access.
| Dataset | Role | Official link |
|---|---|---|
| MIMIC-IV ECG | Pretraining | https://physionet.org/content/mimic-iv-ecg/1.0/ |
| PTB-XL | Probing (4 tasks) | https://physionet.org/content/ptb-xl/1.0.3/ |
| CPSC 2018 / ICBEB | Probing | https://physionet.org/content/challenge-2020/1.0.2/training/cpsc_2018/ |
| Chapman–Shaoxing (CSN) | Probing | https://physionet.org/content/ecg-arrhythmia/1.0.0/ |
| AI-READI | Probing (cardio_relevant) |
https://aireadi.org/dataset · https://fairhub.io/datasets/1 |
After download: see docs/DATA_PREPARATION.md (manifest for MIMIC, folder layout for probing, and config paths).
LVCG pretraining on MIMIC-IV (self-supervised):
python scripts/train.py --config configs/train/lvcg_v5_gru.yamlCheckpoints are written to checkpoints/<run_id>/ (default run id m5s1k1). Use final.pt for downstream evaluation.
python scripts/train.py --config configs/train/lvcg_v5_gru.yaml \
--data.meta_root /path/to/mimic_manifest.jsonlLinear probing on 7 datasets with 10% training labels and seed 42:
python scripts/evaluate.py \
--config configs/eval/probing.yaml \
--checkpoint checkpoints/m5s1k1/final.ptResults are appended to probing/results/probing_results.csv.
Single dataset:
python probing/run_probing.py --config configs/eval/probing.yaml \
--models lvcg --dataset ptbxl_super_class --ratio 0.1 --seed 42lvcg/ # Model, data pipeline, training utilities
scripts/ # train.py, evaluate.py, data helpers
configs/ # train and eval YAML
probing/ # Linear probing + data_splits/
docs/ # DATA_PREPARATION.md, ARCHITECTURE.md
See docs/ARCHITECTURE.md for the LVCG data flow and model design.
If you use this code, please cite the LVCG paper (bibtex TBD).
MIT — see LICENSE.