MITST: Multi-source Irregular Time-Series Transformer for ICU Glucose Level Prediction

MITST (Multi-source Irregular Time-Series Transformer) is a novel machine learning model designed to predict blood glucose (BG) levels in ICU patients using complex, multi-source electronic health record (EHR) data. MITST's modular, hierarchical Transformer architecture processes multi-source and irregular time-series data (e.g., lab results, medications, vital signs), capturing both short-term and long-term temporal dependencies in a flexible way. It offers a scalable and adaptable approach for accurate BG level predictions and other ICU prediction tasks. Below is an illustration of the high-level architecture.

---

Getting Started

To set up and run the code for the MITST model, follow the steps below.

Prerequisites

1. Database Setup
   First, ensure that the eICU database is properly set up as described by the eICU data maintainers. Follow their guidelines at eICU Data Setup.

2. Environment Setup
   Create a virtual environment using Python 3.10. Then, install the required packages by running:

   pip install -r requirements.txt
   

3. Update Patient Table Update the patient table in the database to include previous and next visits. This will be used for concatenating consecutive ICU stays for a single patient within a specific hospital. Run the following command in your SQL environment:

   psql -U your_username -d your_db_name -f concepts/add_previous_next_visits.sql
   
   

Running the Code

Activate your virtual environment and navigate to the code/ directory. Update db_params dictionary in the db_conn.py file with your database information (if necessary):

db_params = {
    "dbname": "your_db_name",
    "user": "your_username",
    "password": "your_password"
}

1. Extract Information from Each Stay Run the following script to extract data from each ICU stay and save it as individual JSON records. Each line in the final JSONL file represents one stay:

   python3 preprocess_database.py
   

2. Concatenate Consecutive Stays and Process Labels Use the following script to concatenate consecutive ICU stays for the same patient. Only stays with at least six BG measurements are included in the final dataset:

   python3 concatenate_admissions_and_preprocess_labels.py
   

3. Feature Generation Now, transform the data to a format compatible with the MITST model by running:

   python3 feature_generation.py
   

4. Train the MITST Model Once the data is prepared, train the MITST model using:

   python3 train.py
   

5. Model Evaluation Finally, after training, evaluate the model on the test set to compute AUROC and AUPRC metrics for each class:

   python3 evaluation.py
   

Citation

If you find MITST helpful in your research, please consider citing our preprint:

@article{mehdizavareh2024mitst,
  title={Enhancing Glucose Level Prediction of ICU Patients through Irregular Time-Series Analysis and Integrated Representation},
  author={Mehdizavareh, Hadi and Khan, Arijit and Cichosz, Simon Lebech},
  journal={arXiv preprint arXiv:2411.01418},
  year={2024}
}

Acknowledgements

1. Special thanks to the eICU team for providing access to this large-scale EHR dataset. Their work enables impactful research in healthcare analytics. Learn more at eICU GitHub Repository.
2. We extend our appreciation to the teams behind the FT-Transformer and Tab-Transformer for their open-source implementations, which inspired aspects of MITST. See their project at FT-Transformer GitHub Repository.