Add ERReadmissionMIMIC4 task for ER-specific 30-day readmission prediction

[RajendranRanjith]

Contributor Information

• Name: Ranjithkumar Rajendran
• NetID: rr54
• Email: rr54@illinois.edu
• Course: CS598 DL4H — Spring 2026

Contribution Type

Standalone Task — New task for MIMIC-IV ER-specific readmission prediction.

Paper Reference

KEEP: Integrating Medical Ontologies with Clinical Data for Robust Code Embeddings

• Authors: Elhussein et al.
• Venue: CHIL 2025
• Link: https://arxiv.org/abs/2510.05049

Description

This PR adds ERReadmissionMIMIC4, a new PyHealth task that predicts 30-day readmission for Emergency Room patients in MIMIC-IV. This implements the ER cohort definition from the KEEP paper, which demonstrated that ER patients have distinct readmission patterns requiring specialized prediction models.

How it works:

1. Filters admissions where admission_location == 'EMERGENCY ROOM'
2. Extracts versioned ICD diagnosis codes (e.g., "10_E11.9")
3. Computes discharge time from the admissions table
4. Labels each visit: 1 if next admission < 30 days, 0 otherwise
5. Returns PyHealth-compatible sample dicts with conditions and readmission fields

Key design decisions:

• Inherits from BaseTask following PyHealth conventions
• Configurable readmission window via timedelta parameter (default: 30 days)
• Skips patients with < 2 admissions and visits with no diagnoses
• ICD codes versioned as "{version}_{code}" matching ReadmissionPredictionMIMIC4
• PEP8 compliant (88-char limit), Google-style docstrings, full type hints

File Guide

#	File	Description
1	pyhealth/tasks/mimic4_er_readmission.py	Core task implementation — ERReadmissionMIMIC4 class (~130 lines)
2	pyhealth/tasks/__init__.py	Updated to export ERReadmissionMIMIC4
3	tests/test_mimic4_er_readmission.py	7 unit tests using synthetic data (edge cases: single admission, no diagnoses, custom window, ICD versioning)
4	examples/mimic4_er_readmission_retain.py	Ablation script: Standard vs ER-specific readmission using RETAIN
5	examples/mimic4_er_readmission_transformer.py	Ablation script: RETAIN vs Transformer on ER cohort
6	docs/api/tasks/pyhealth.tasks.mimic4_er_readmission.rst	Sphinx API documentation
7	docs/api/tasks.rst	Updated task index with new entry

Tests

All 7 unit tests pass using synthetic Patient objects (no real data required):