Physical AI — Digital Twin + AI Hybrid Framework for HFpEF

Physical AI is an integrated framework that combines digital twin cardiovascular models with modern machine learning / generative AI techniques.
This repository is the starting platform for future development of full digital twin–guided physiological AI applications.

Currently, the repo includes two core modules:

1. Physically-Informed Generative AI (phys_informed_gen_ai)
2. Classifiers Learning from Digital Twin Simulations (classifiers_learn_sims)

Both modules live under src/ and interact with data, pretrained models, digital twin simulations, and output results.

---

📁 Repository Structure

PhysicalAI/
│
├── data/                         # All datasets used by the two modules
│
├── model/                        # Trained VAE models & classifier PyTorch checkpoints
│
├── result/                       # Synthetic data, evaluation plots, predictions
│
└── src/
    ├── phys_informed_gen_ai/     # VAE + digital twin hybrid generative module
    │   ├── main.ipynb            # Main notebook for VAE-based generative DT framework
    │   └── *.m                   # MATLAB scripts for physical constraints (DT validation)
    │
    └── classifiers_learn_sims/   # Classifier training & prediction module
        └── main.ipynb            # Main notebook for training & inference

---

1. Physically-Informed Generative AI (phys_informed_gen_ai)

This module uses a VAE trained on HFpEF digital twin phenotypes to generate new virtual digital twin samples.

Three Sampling Modes

(1) VirtualDT_original

• Random sampling from the VAE latent space
• No covariance constraints

(2) VirtualDT_correlated

• Sampling from a multivariate Gaussian that preserves the training-set latent covariance
• Produces more realistic synthetic phenotypes

(3) VirtualDT_phys

• Based on (2), but each generated sample is validated using the digital twin model (MATLAB)
• Only physiologically consistent samples are retained
• This forms a hybrid digital twin + AI dataset

Generated data are stored under:

result/
  ├── VirtualDT_original.csv
  ├── VirtualDT_correlated.csv
  └── VirtualDT_phys.csv

📊 Key Plots

Located in result/:

• Z_Sampling_Methods_Comparison.png
• VAE_reconstruction_PDF_CC.png
• VAE_reconstruction_PDF_CC.png

These compare the latent sampling methods and reconstruction properties.

🧠 Digital Twin Validation

All MATLAB scripts under this directory are used to:

• Evaluate generated samples for physiological plausibility
• Reject unrealistic digital twins
• Produce final “physically-informed” synthetic datasets

---

2. Classifiers Learning from Simulations (classifiers_learn_sims)

This module trains classification models using labels derived from digital twin simulations.

📄 Training & Testing Data

Located in data/:

• TrainingC.csv
• TestingC.csv

🧠 Trained Classifier Models

Located in model/:

• model_LabelCO_best.pth
• model_LabelFillingP_best.pth
• (additional best models)

📊 Evaluation Results

In result/:

evaluation_plots/           # Confusion matrices (train/validation)
external_validation_plots/  # Confusion matrices for external test set

---

3. Predicting New Patient Data (Most Directly Usable Feature)

You can run the full pipeline on new clinical data.

Step 1 — Prepare New Patient File

Replace the placeholder file:

data/inputclassifierMYPACE.csv

The repo includes 10 synthetic example patients.
Replace them with real patients using identical column names.

Step 2 — Run the Prediction Cell

Inside:

src/classifiers_learn_sims/main.ipynb

Run the second-to-last cell (the script clearly marks it).

Step 3 — View Output Predictions

Results will appear in:

result/inputclassifierMYPACE_with_predictions.csv

Step 4 — More Details

The final cell in main.py contains a detailed explanation of the inference pipeline.

---

🌟 Future Development

This repository is the initial version of a broader “Physical AI” framework which will include:

• PINN-based digital twin inversion
• Longitudinal prediction models
• Adaptive pacing-response estimation
• Full digital-heart physical simulation integration

Future updates will expand the src/ modules and add additional datasets and models.

---

📦 Requirements

• Python 3.8–3.11
• PyTorch
• NumPy / SciPy / Pandas / Matplotlib
• MATLAB + MATLAB Engine for physically-informed VAE filtering

---

🧑‍💻 Contact

For questions or collaboration:

Feng Gu
University of Michigan
Email: fenggu@umich.edu