Hybrid PINN with Experimental Data

Reproducibility code for: Hybrid PINNs with Experimental Data for Real-Time Hemodynamic Prediction and Clinical Decision Support — International Journal of Medical and Pharmaceutical Research, December 2025.

Paper summary

Pure physics-informed networks rely on the assumption that the modeled equations capture every relevant phenomenon. In real arteries this is rarely true: vessel wall compliance, branch flow splits, and patient-specific viscosity all introduce model error. This paper presents a hybrid framework in which sparse experimental measurements (4D-flow MRI velocity samples and Doppler ultrasound peak velocities) are used as a soft data constraint alongside the Navier–Stokes residual.

The contribution is a principled weighting scheme between physics and data losses based on the local measurement uncertainty, allowing the network to defer to data where measurements are dense and to physics where they are sparse. We demonstrate sub-second inference on retrained models, supporting use as real-time clinical decision support during interventional procedures.

What this repository contains

pinn-hybrid-experimental/
├── src/
│   ├── data_assimilation.py    # Uncertainty-weighted data loss
│   └── realtime_inference.py   # Quantized inference for sub-second eval
├── configs/                    # (empty)
├── experiments/                # (empty)
├── notebooks/                  # (empty)
└── scripts/                    # (empty)

Method

The training loss extends the standard PINN loss with a measurement-fitting term:

L = λ_pde · L_NS  +  λ_bc · L_BC  +  Σ_k (1 / σ_k²) · ||u_θ(x_k) − ū_k||²

where (x_k, ū_k, σ_k) is the k-th measurement (location, value, standard deviation). The inverse-variance weighting matches the Bayesian interpretation of measurements as Gaussian observations of the true field, and removes the need to hand-tune a single global λ_data.

Status

This repository contains the method components for the paper's hybrid formulation, not a runnable reproduction pipeline.

Component	State
Uncertainty-weighted data-assimilation loss (src/data_assimilation.py)	Implemented
Real-time inference wrapper (src/realtime_inference.py)	Implemented
Measurement loaders (4D-flow MRI, Doppler)	Not implemented
Training and benchmark scripts	Not implemented
Configs, notebooks, figures	Not present

The core PINN implementation lives in VascuPINN, which is itself a reference implementation rather than the code used for the published results.

git clone https://github.com/avcodes438/pinn-hybrid-experimental.git
cd pinn-hybrid-experimental
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Data availability

No measurement data is released in this repository. The clinical measurements referenced in the paper are not publicly available here, and no dataset release is currently published. Anyone seeking the underlying data should contact the author directly.

Citation

@article{vaghela2025hybrid,
  title   = {Hybrid PINNs with Experimental Data for Real-Time Hemodynamic Prediction and Clinical Decision Support},
  author  = {Vaghela, Aaryasinh},
  journal = {International Journal of Medical and Pharmaceutical Research},
  year    = {2025},
  month   = {12}
}

License

MIT — see LICENSE.

Contact

Aaryasinh Vaghela — aaryasinh.vaghela@gmail.com