📊 Glycemic Control

Modeling causal relationships between glucose levels and clinical interventions in ICU settings using graph-based deep learning.

Overview

Standard predictive models in critical care capture correlations without explaining why changes occur. This project uses causal modeling with Graph Neural Networks to identify variables influencing glucose regulation.

Key approaches:

• Causal discovery to build structural models connecting clinical variables to glucose outcomes
• Graph Neural Networks (GNNs) with DAG representations for learning relationships
• Interpretable insights from high-dimensional ICU time series data

Project Structure

glycemic_control/
├── src/
│   ├── constants.py           # Clinical variable definitions
│   ├── dataset.py             # PyTorch datasets
│   ├── dataloader.py          # GNN data loading
│   ├── utils.py               # Utilities (early stopping, LR scheduler)
│   ├── train_gnn.py           # GNN training pipeline
│   ├── train_gnn_set.py       # Batch training across DAGs
│   └── models/
│       ├── temporal_gnn.py    # DCRNN-based models
│       └── transformer.py     # Time series transformer
├── notebooks/                 # Exploration notebooks
├── dataset_scripts.py         # Data preprocessing
├── clinical_dag_scripts.py    # Clinical DAG processing
├── aggregate_dags_script.py   # DAG aggregation via majority voting
└── requirements.txt

Installation

conda create -n glycemic python=3.8
conda activate glycemic
pip install -r requirements.txt

Key dependencies: PyTorch, PyTorch Geometric, PyTorch Geometric Temporal, pandas, scikit-learn

Clinical Variables

Category	Variables
Organ Scores	SOFA, liver, CNS, cardiovascular, coagulation, respiration, renal
Patient Metrics	weight, height, BMI, age, Charlson index
Interventions	tube feeding, dextrose, insulin
Target	glucose (discretized into 5 classes)

Usage

1. Prepare Dataset

python dataset_scripts.py --fpath_raw_csv /path/to/cohort.csv --dirpath_out /path/to/output/

2. Process Clinical DAGs

python clinical_dag_scripts.py --fpath_excel /path/to/dags.xlsx --dag_type diabetic --dirpath_save /path/to/dags/

3. Train GNN Model

python -m src.train_gnn \
    --fpath_train_data train.csv \
    --fpath_valid_data valid.csv \
    --fpath_test_data test.csv \
    --fpath_dag dag.csv \
    --dirpath_results /path/to/results/

Models

• RecurrentGCN: DCRNN-based graph neural network for temporal classification
• TimeSeriesTransformer: Encoder-decoder transformer for sequence forecasting

References

• Wu et al. (2020). Deep Transformer Models for Time Series Forecasting
• Li et al. (2018). Diffusion Convolutional Recurrent Neural Network

---