Predicting Circulatory Failure in ICU Patients with Deep Sequential Models

This repository archives the implementation associated with the paper Predicting Circulatory Failure in ICU Patients with Deep Sequential Models, presented at CBIC 2025.

• Event page: https://sbia.org.br/eventos/cbic_2025/cbic2025-1191697/
• DOI: https://doi.org/10.21528/CBIC2025-1191697
• Dataset: MIMIC-IV
• Prediction horizon: up to 8 hours before circulatory failure
• Temporal resolution: 5 minutes

The expanded multi-cohort work for the journal manuscript is maintained in lucaspimentab/multicohort-circulatory-failure.

Models

The study compares three deep sequential architectures:

• Transformer
• LSTM
• GRU

Pretrained weights for three random seeds per architecture are included under models/.

Repository Structure

data/
  features_names.csv       # model feature names
  metrics.csv              # aggregate evaluation metrics
figures/                   # publication figures
models/                    # pretrained model weights
notebooks/
  extract_stayids_and_split.ipynb
  preprocess.ipynb
  models.ipynb
  graphs.ipynb
environment.yml
requirements.txt

Patient identifiers, raw MIMIC-IV data, and generated patient-level datasets are not distributed in this repository.

Setup

Using Conda:

conda env create -f environment.yml
conda activate circulatory-failure

Or using pip:

pip install -r requirements.txt

Copy .env.example to .env and configure the local MIMIC-IV PostgreSQL connection. Credentials must never be committed.

Workflow

1. Run notebooks/extract_stayids_and_split.ipynb to create patient-level train, validation, and test splits locally.
2. Run notebooks/preprocess.ipynb to build the model-ready HDF5 dataset.
3. Run notebooks/models.ipynb to train and evaluate the sequence models.
4. Run notebooks/graphs.ipynb to generate the evaluation figures.

Reported Results

Model	AUROC	AUPRC	F1	Precision	Recall
Transformer	0.976	0.884	0.804	0.844	0.768
LSTM	0.969	0.858	0.745	0.839	0.671
GRU	0.969	0.839	0.782	0.812	0.755

Data Governance

MIMIC-IV is not distributed with this repository. Access requires PhysioNet credentialing and compliance with the applicable data use agreement. Keep all credentials, raw data, patient identifiers, and patient-level outputs outside Git.