Intensive Documentation Index (IDI) — MIMIC-IV Development & Validation

Paper: Development and Validation of the Intensive Documentation Index for ICU Mortality Prediction
Journal: Journal of the American Medical Informatics Association (JAMIA), 2026
Authors: Alexis M. Collier, DHA, MHA¹ · Sophia Z. Shalhout, PhD²³
Affiliations:
¹ College of Health & Wellness, University of North Georgia, Dahlonega, GA, USA
² Department of Otolaryngology–Head and Neck Surgery, Harvard Medical School, Boston, MA, USA
³ Mass Eye and Ear, Mass General Brigham, Boston, MA, USA

Companion paper (multinational validation): IDI-Multinational-Validation (npj Digital Medicine, 2026)

---

Overview

This repository contains the full analysis code for the Intensive Documentation Index (IDI) — a zero-burden prognostic framework that extracts temporal documentation rhythm features from nursing chartevents timestamps in the first 24 hours of ICU admission to predict in-hospital mortality.

Applied to 26,153 heart failure ICU admissions from MIMIC-IV (2008–2019), the IDI modestly but reliably improves mortality prediction beyond traditional clinical variables.

---

Key Results

Model	AUROC (95% CI)	Calibration Slope	Brier Score
Baseline (age, sex, ICU LOS)	0.658 (0.609–0.710)	0.92	0.1091
IDI-Enhanced	0.683 (0.631–0.732)	0.96	0.1080

ΔAUROC = +0.025 (p = 0.015, DeLong test)
Strongest predictor: idi_cv_interevent OR = 1.53 per SD (95% CI 1.35–1.74, p < 0.001)
Temporal stability: mean AUC 0.654 (SD 0.016) across leave-one-year-out cross-validation (2008–2019)

---

Repository Structure

IDI-MIMIC-IV-Mortality/
│
├── README.md
├── LICENSE
├── requirements.txt
├── .gitignore
│
├── src/
│   ├── idi_features.py        # IDI feature extraction from chartevents timestamps
│   ├── cohort_selection.py    # Inclusion/exclusion criteria applied to MIMIC-IV
│   ├── model.py               # Logistic regression training & temporal validation
│   ├── metrics.py             # AUROC, calibration slope, Brier, DeLong test
│   └── equity_analysis.py     # Subgroup AUC analysis by race/ethnicity
│
├── data/
│   └── README.md              # Data access instructions (PhysioNet DUA required)
│
└── results/
    ├── figures/               # ROC curves, calibration plots, forest plots
    └── tables/                # CSV versions of manuscript tables

---

Data Access

Raw data are NOT included in this repository due to PhysioNet Data Use Agreement restrictions.

To reproduce this analysis:

1. Apply for credentialed access at PhysioNet
2. Download MIMIC-IV version 2.2: https://physionet.org/content/mimiciv/2.2/
3. Place the following files in data/mimic-iv/:
   • hosp/admissions.csv
   • hosp/patients.csv
   • hosp/diagnoses_icd.csv
   • icu/icustays.csv
   • icu/chartevents.csv

---

Installation

git clone https://github.com/colla00/IDI-MIMIC-IV-Mortality.git
cd IDI-MIMIC-IV-Mortality
pip install -r requirements.txt

Python version: 3.8+

---

Usage

Run scripts in order:

# 1. Select cohort (outputs data/cohort.csv)
python src/cohort_selection.py

# 2. Extract IDI features (outputs data/idi_features.csv)
python src/idi_features.py

# 3. Train models and run temporal validation (outputs results/)
python src/model.py

# 4. Compute performance metrics
python src/metrics.py

# 5. Run equity analysis (outputs results/figures/equity_forest.png)
python src/equity_analysis.py

---

IDI Features

Nine temporal features extracted from nursing chartevents in the first 24 ICU hours:

Feature	Domain	Description
idi_events_24h	Volume	Total documentation events
idi_events_per_hour	Volume	Event rate per hour
idi_max_gap_min	Surveillance Gap	Maximum inter-event interval (min)
idi_gap_count_60m	Surveillance Gap	Intervals > 60 minutes
idi_gap_count_120m	Surveillance Gap	Intervals > 120 minutes
idi_mean_interevent_min	Rhythm	Mean inter-event interval (min)
idi_std_interevent_min	Rhythm	SD of inter-event intervals
idi_cv_interevent	Rhythm	Coefficient of variation (SD/mean)
idi_burstiness	Rhythm	Burstiness index B = (σ−μ)/(σ+μ)

---

Temporal Leakage Prevention

Features with absolute Pearson correlation > 0.30 with ICU length of stay were removed prior to modeling to prevent reverse-causal leakage (longer stay → more documentation → spurious mortality prediction). See Methods section of the manuscript for full details.

---

Citation

If you use this code, please cite:

@article{collier2026idi,
  title     = {Development and Validation of the Intensive Documentation Index for ICU Mortality Prediction},
  author    = {Collier, Alexis M. and Shalhout, Sophia Z.},
  journal   = {Journal of the American Medical Informatics Association},
  year      = {2026},
  doi       = {[to be assigned]}
}

---

Patent Notice

The IDI framework is the subject of U.S. provisional patent applications (Patent Pending):

• USPTO Application No. 63/976,293 — System and Method for Predicting ICU Mortality from Electronic Health Record Documentation Rhythm Patterns (filed February 2026)
• USPTO Application No. 63/946,187 — Clinical Decision Support System with Trust-Based Alert Prioritization and Equity Monitoring (filed December 2025)

VitaSignal LLC is the intended assignee. Licensing inquiries: info@vitasignal.ai

---

Funding

This research was, in part, funded by the National Institutes of Health (NIH) Agreement No. 1OT2OD032581 through the AIM-AHEAD program.

---

License

MIT License — see LICENSE file for details.