Add MIMIC-III circulatory failure dataset and prediction task with ablation study

[kywang0906]

Summary

This PR adds a new dataset and task for early circulatory failure prediction on MIMIC-III, inspired by the FAMEWS framework for fairness-aware early warning systems.

The contribution includes:

• A custom dataset (MIMIC3CirculatoryFailureDataset) for cohort construction and MAP time-series extraction
• A task (CirculatoryFailurePredictionTask) for generating time-point prediction samples
• Example scripts demonstrating task usage and ablation studies
• Unit tests using synthetic data

---

Paper Reference

This work is inspired by:

FAMEWS: Fairness Auditing for Medical Early-Warning Systems

The task design follows the early-warning prediction setting described in the paper, focusing on predicting circulatory failure events within a future time window.

---

Contributors:

Kuang-Yu Wang (Net Id: kuangyu4, Email: kuangyu4@illinois.edu)
Ya Hsuan Yang (Net Id: yhyang3, Email: yhyang3@illinois.edu)

---

Contribution Type

• New Dataset
• New Task
• New Model

---

File Overview

Dataset

• pyhealth/datasets/mimic3_cf.py
  Implements cohort construction, failure label generation, and MAP time-series extraction

• pyhealth/datasets/configs/mimic3_cf.yaml
  Dataset configuration for MIMIC-III tables

Task

• pyhealth/tasks/circulatory_failure_prediction.py
  Defines input/output schema and label generation logic

Example

• examples/mimic3_cf_circulatory_failure_logreg.py
  Demonstrates dataset-task pipeline and ablation study

Tests

• tests/core/test_circulatory_failure_prediction.py
• tests/core/test_mimic3_cf.py

---

Ablation Study

We conduct ablation experiments on:

1. Prediction Window

• 6 hours
• 12 hours
• 24 hours

Results show that shorter prediction windows yield stronger predictive signals, while longer horizons introduce uncertainty and reduce performance.

2. Feature Design

• Baseline: MAP only
• Advanced: MAP + temporal difference (map_diff)

The advanced feature setting improves recall by capturing temporal trends in physiological signals.

3. Fairness Analysis (Subgroup Evaluation)

We evaluate performance across gender subgroups (M/F) and observe substantial disparities in recall and ROC-AUC, consistent with findings in FAMEWS regarding fairness issues in early warning systems.

---

Testing

All tests are implemented using synthetic data to ensure fast execution:

• Task tests validate label generation logic and output format
• Dataset tests validate integration with task and sample generation

---

Notes

• Due to MIMIC-III de-identification, timestamps are shifted; only relative temporal relationships are used.

[jhnwu3]

Will review further, here, but this is a big issue with the PR.

[jhnwu3]

Hey,I don't think you need to wrap MIMIC3 again, see the existing implementation here for MIMIC3. I don't see how this is different from our MIMIC3 implementation.