Cardiac Digital Twin — PIGNN for ECG Synthesis
HardJun 2025 — Present
Physics-informed graph neural network that models the heart's electrical conduction system and synthesizes 12-lead ECG from PPG input.
What it does
- Modeled the SA node, AV node, His bundle, and Purkinje network as a graph using Graph GRU message-passing to synthesize 12-lead ECG from pulse-oximetry input.
- Combined a rule-based HRV classifier (bradycardia, tachycardia, arrhythmia, hypoxemia) with the deep model, trained on PhysioNet's MIT-BIH and PTB databases.
- Built a 1D convolutional encoder + temporal-refinement GRU with a composite loss (MSE, Pearson correlation, physiological regularization).
- Shipped a PyTorch pipeline with checkpointing, patient-level splitting, and cross-platform fixes.
Tags
PythonPyTorchGraph Neural NetworksJupyter
solution.md
# stack
- Python
- PyTorch
- Graph Neural Networks
- Jupyter
# status
- attempted — code public, no hosted demo yet