RADAR

RADAR (Risk Assessment via Decay-weighted Aggregation of Records) predicts the risk of future incident atrial fibrillation (AF) from longitudinal diagnosis histories. It pools each encounter's diagnosis set, updates a set of memory embeddings with time-dependent decay, and estimates monthly AF and death hazards through one year.

This repository provides the model, training and prediction scripts, a MIMIC-IV adapter, and a synthetic demo for Deep Learning for Population-Scale Prediction of Incident Atrial Fibrillation. The paper configuration is in configs/paper.json.

Installation

Requires Python 3.10 or newer and PyTorch 2.2 or newer.

python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[test]"
pytest

For CUDA, install the appropriate PyTorch build before installing this package.

Demo

bash scripts/run_demo.sh

This produces synthetic data following a similar format as available within MIMIC, prepares patient sequences, trains a small model for one epoch, and writes demo_run/predictions/predictions.csv. No external data are needed.

MIMIC-IV

Obtain MIMIC-IV v3.1 through PhysioNet after completing its credentialing and data-use requirements. The adapter reads patients, admissions, diagnoses_icd, and procedures_icd CSV files (plain or gzip-compressed) under /path/to/mimic-iv-3.1/hosp/.

Separately download the 2018 ICD-10-CM GEM ZIP from the official CMS archive and save it locally as data/2018_I10gem.zip. The adapter reads the forward ICD-10-CM → ICD-9-CM diagnosis mapping (2018_I10gem.txt) directly from the ZIP. Use the diagnosis GEM, not the reverse or procedure mapping.

mkdir -p data
# Save the CMS download as data/2018_I10gem.zip before preparing data.
radar-prepare-mimic \
  --mimic-dir /path/to/mimic-iv-3.1 \
  --gem data/2018_I10gem.zip \
  --out-dir prepared/mimic-radar

radar-train \
  --data-dir prepared/mimic-radar \
  --out-dir runs/mimic-radar \
  --epochs 5

radar-predict \
  --checkpoint runs/mimic-radar/checkpoint_best.pt \
  --data-dir prepared/mimic-radar \
  --split test \
  --out-dir outputs/mimic-radar-test

Training uses AdamW with 5% linear learning-rate warm-up followed by cosine decay. The best checkpoint maximizes full 12-month validation AUROC among encounters with observable binary outcomes.

Preparation uses patient-level 70/15/15 splits and learns the vocabulary and age normalization from training patients only. AF outcome codes are removed from model inputs; the default postoperative exclusion window is 90 days. Predictions include diagnoses assigned during the current admission and are retrospective encounter-level estimates. Follow-up is capped at 365 days, with death treated as a competing event.

These commands train a MIMIC model using vocabulary and normalization learned from its training patients. The paper's external validation applies the frozen HCAI-trained model with its HCAI vocabulary and normalization. Dataset access and derived outputs are governed by the provider's data-use terms. Run each command with --help for options.

Other datasets

Use the runnable custom-data example to adapt RADAR to another encounter-level dataset. The data schema describes the prepared arrays; the model card summarizes inputs, outputs, and intended use.

Citation

If you use RADAR, please cite the accompanying manuscript:

Alexander Schubert, Abhijith Varma Mudunuri, Marie-Laure Charpignon, Jean Jacques Noubiap, Ahmed Alaa, Gregory M. Marcus. Deep Learning for Population-Scale Prediction of Incident Atrial Fibrillation. Under review.

Citation details will be updated upon publication.

Please also cite MIMIC-IV when using its data:

Johnson AEW et al. MIMIC-IV, a freely accessible electronic health record dataset. Scientific Data. 2023;10:1.

MIMIC-IV v3.1: https://physionet.org/content/mimiciv/3.1/

License

Code is released under the MIT License. Dataset access and use are governed separately by the relevant data providers.