Graph-spa: A Spatiotemporal Graph Neural Network based framework for ARDS Prediction and Interpretability

This repository contains code and experiments for training and evaluating a spatiotemporal graph neural network model on three large-scale ICU datasets: HiRID, MIMIC-IV, and eICU-CRD. The work aims to predict acute ARDS onset and supports interpretability through attribution mask generation.

---

📁 Repository Structure

/training

Notebooks and scripts for training baseline and denoising autoencoder (DAE) variants of the model on each dataset.

• 01a_HiRID_baseline.ipynb, 01b_HiRID_dae.ipynb: HiRID training
• 02a_MIMIC_baseline.ipynb, 02b_MIMIC_dae.ipynb: MIMIC-IV training
• 03a_EICU_baseline.ipynb, 03b_EICU_dae.ipynb: eICU training
• 04_extval_MIMIC_EICU.ipynb, 05_extval_EICU_MIMIC.ipynb: External validation notebooks
• layer.py, net.py, loader.py: baseline model architecture and data loading
• layer_dev.py, net_dev.py: Dynamic adjacency version
• mimicandeicufeatures.txt: Feature list used for MIMIC and eICU processing

---

/results

Notebooks to generate all figures and statistical analyses reported in the paper.

📊 Main Figures

• 01_Fig2a_plot_hirid.ipynb, 02_Fig2b_plot_mimic4.ipynb, 03_Fig2c_plot_eICU.ipynb: Dataset-specific performance plots
• 04_Fig3a_adjacencyplots.ipynb, 04_Fig3bc_plot.ipynb: Adjacency Comparison
• 05_Ablation_statistical_analysis.ipynb: Statistical comparisons for ablation experiments
• 06_Fig4ab_Attribution_Masks.ipynb, 06_Figure4cde_HiRID_attribution_plot.ipynb: Attribution mask visualizations
• 07_Fig05a_HiRID.ipynb, 07_Fig05b_MIMIC.ipynb, 07_Fig05c_eICU.ipynb: Co-occurrence analysis.

🧪 Extended Figures and Supplement

• 08_FigS01a_plot_mimic_eicu.ipynb, 09_FigS01b_plot_eicu_mimic.ipynb: Supplementary external validation plots
• 10_FigS02_mimic_eicu_attribution_plot.ipynb, 11_MIMIC_and_eICU_Attribution_Masks.ipynb: Additional attribution mask insights

---

📦 Data Access Instructions

---

🔁 Results Reproducibility

To reproduce all result figures in /results:

1. Download the precomputed result files from Zenodo:
   👉 https://zenodo.org/records/15924818
2. Extract the archive into the results/ folder of this repository.
3. Run the jupyter notebooks to generate the plots.

---

🧪 Training Data Preparation

To run the /training notebooks you need the raw ICU time‑series:

• HiRID, MIMIC‑IV, eICU‑CRD
  Available on PhysioNet https://physionet.org/ (credentialed access required).
• MIMIC‑IV & eICU‑CRD
  After download, preprocess with the METRE pipeline https://github.com/weiliao97/METRE to extract and normalize features.

---

📌 Dependencies

This repository is implemented using Python 3.11.5 and requires the packages in requirements.txt:

Install all requirements using:

pip install -r requirements.txt

🔍 Contact

For questions or collaboration inquiries, please contact:

Shashank Yadav
PhD Candidate, University of Arizona
Email: [shashank@arizona.edu]