🧄 GARLIC: Graph Attention-Based Relational Learning of Multivariate Time Series in Intensive Care

This repository contains the official PyTorch implementation of GARLIC, accepted at ICLR 2026.

GARLIC is a novel neural framework designed for accurate and interpretable clinical outcome prediction (e.g., mortality, sepsis) from irregularly sampled, multivariate Intensive Care Unit (ICU) time series. It achieves state-of-the-art performance on major benchmarks (PhysioNet-12, PhysioNet-19, and MIMIC-III) while providing transparent, built-in explanations at the observation, signal, and inter-signal relational levels.

🌟 Overview

Clinical data is often messy, irregular, and riddled with missing values. GARLIC tackles these challenges without sacrificing interpretability through a three-stage architecture:

1. Latent Feature Modeling: Handles irregular missingness using a learnable exponential-decay encoder.
2. Time-Lagged Graph Message Passing: Captures dynamic inter-sensor dependencies through learned summary graphs.
3. Cross-Dimensional Sequential Attention: Fuses global patterns across time and signals for robust prediction.

Figure: The GARLIC Architecture.

⚙️ Requirements

The code is tested with Python 3.11. We recommend setting up a virtual environment (e.g., conda or venv) before installing dependencies.

Install the required packages using pip:

pip install -r requirements.txt

📊 Datasets & Preparation

We evaluate GARLIC on three standard public ICU datasets. You will need to download the raw data and place them in the correct directories.

Download Links

• PhysioNet Challenge 2012 (P12): Download here
• PhysioNet Challenge 2019 (P19): Download here
• MIMIC-III: Request access here (Credentialed access required)

Directory Structure

Place the downloaded raw datasets into the following structure:

├── data/
│   ├── rawdata/
│   │   ├── P12/           # Place P12 raw files here
│   │   ├── P19/           # Place P19 raw files here
│   │   └── MIMICIII/      # Place MIMIC-III raw files here
│   └── processed_data/    # Auto-generated during the first run
├── run.sh
├── interpretability_evaluation.sh
└── ...

Note: Preprocessing will automatically trigger during the first run and save the cleaned data to ./data/processed_data/.

🚀 Running the Model

1. Training & Evaluation To run the full pipeline (data preprocessing, model initialization, training, and evaluation), execute the main shell script. This will output the AUROC and AUPRC metrics for the tasks.

bash run.sh

2. Interpretability Evaluation To reproduce the quantitative interpretability experiments (e.g., the perturbation-based masking using Top 50%, Bottom 50%, Random 50%), run:

bash interpretability_evaluation.sh

🧪 Reproducibility notes

• Fix random seeds (if exposed via flags/configs) for comparable results.
• Small numerical differences may occur across hardware and CUDA versions.
• The first run may take longer due to preprocessing and caching under ./data/processed_data/.

🙏 Acknowledgments

This research project was partially supported by the Schweizer Paraplegiker Stiftung and the ETH Zürich Foundation (2021-HS-348) and the JST Moonshot R&D Program, Grant Number JPMJMS2034-18.

📖 Citation

If you find this code or our paper useful for your research, please consider citing:

@inproceedings{wang2026garlic,
  title={{GARLIC}: Graph Attention-Based Relational Learning of Multivariate Time Series in Intensive Care},
  author={Wang, Ruirui* and Li, Yanke* and G{\"u}nther, Manuel and Paez-Granados, Diego},
  booktitle={The Fourteenth International Conference on Learning Representations (ICLR)},
  year={2026},
  url={[https://openreview.net/forum?id=4ZAwmIaA9y](https://openreview.net/forum?id=4ZAwmIaA9y)}
}