FLAME-ICU

Federated Learning Adaptable Mortality Estimator for the ICU

A multi-site federated learning implementation of the FLAIR benchmark for ICU outcome prediction using CLIF-standardized data.

---

What is FLAIR?

Note: FLAIR is currently in development and will soon be available via PyPI as flair-benchmark.

FLAIR (Federated Learning for AI Research) is a privacy-preserving benchmark framework for ICU prediction tasks. It provides:

• 7 clinically relevant prediction tasks for ICU outcomes
• Standardized data format using CLIF (Common Longitudinal ICU Format)
• Privacy-preserving design - no data leaves your institution
• Reproducible benchmarks across 17+ US hospitals

FLAIR Tasks

Task	Name	Type	Prediction Window
5	ICU Length of Stay	Regression	24 hours
6	Hospital Mortality	Classification	24 hours
7	ICU Readmission	Classification	Entire ICU stay

FLAME-ICU implements Tasks 5, 6, and 7.

---

Quick Start

1. Clone FLAME-ICU

git clone https://github.com/Common-Longitudinal-ICU-data-Format/FLAME-ICU.git
cd FLAME-ICU

2. Clone FLAIR inside the FLAME repo

git clone https://github.com/Common-Longitudinal-ICU-data-Format/FLAIR.git

3. Sync and install

uv sync
uv pip install -e ./FLAIR

4. Download Rush models

Download the rush_models/ folder from the CLIF BOX FLAME folder (link shared privately) and place it in the FLAME-ICU root directory.

5. Configure

Make sure your clif_config.json file is properly renamed and updated with site paths and timezone.

Run the pipeline

# Step 1: Generate datasets
uv run marimo edit code/notebooks/01_task_generator.py

# Step 2: Feature engineering
uv run marimo edit code/notebooks/02_feature_engineering.py

# Step 2b: Generate Table 1
uv run python code/notebooks/02b_table_one.py

# Step 3: Train & evaluate (run all three)
uv run marimo edit code/notebooks/03a_task5_icu_los.py
uv run marimo edit code/notebooks/03b_task6_mortality.py
uv run marimo edit code/notebooks/03c_task7_readmission.py

See RUN.md for details.

---

Project Structure

FLAME-ICU/
├── FLAIR/                    # FLAIR benchmark library
│   └── flair/                # Core library code
├── code/
│   ├── notebooks/            # Marimo notebooks (main workflow)
│   │   ├── 01_task_generator.py      # Generate task datasets
│   │   ├── 02_feature_engineering.py # Extract CLIF features
│   │   ├── 03a_task5_icu_los.py      # ICU LOS models
│   │   ├── 03b_task6_mortality.py    # Mortality models
│   │   └── 03c_task7_readmission.py  # Readmission models
│   ├── models/               # Model implementations
│   │   ├── xgboost_model.py
│   │   ├── elasticnet_model.py
│   │   └── evaluation.py
│   └── features.py           # Feature extraction
├── outputs/                  # Generated datasets and features
├── rush_models/              # Rush baseline models
├── results_to_box/           # Site results for upload
├── optimization/             # Hyperparameter tuning (Optuna)
├── clif_config.json          # Site configuration
├── RUN.md                    # Execution guide
└── README.md                 # This file

---

Federated Learning Workflow

Rush Site (Baseline)

1. Train XGBoost and ElasticNet models on Rush data
2. Upload trained models to BOX
3. Models serve as baseline for other sites

Other Sites (Evaluation)

1. Download Rush models from BOX
2. Run 4 experiments per task:
   • Rush Evaluation - Test Rush models on local data
   • Platt Scaling - Calibrate Rush models with local train data
   • Transfer Learning - Fine-tune Rush models with local data
   • Independent - Train from scratch on local data
3. Upload results to BOX

---

Models

XGBoost

• Gradient boosting with native missing value handling
• Optimized hyperparameters via Optuna (50 trials, 3-fold CV)

ElasticNet

• L1+L2 regularized linear model
• StandardScaler + median imputation
• LogisticRegression for classification, ElasticNet for regression

---

Required CLIF Tables

Table	Key Columns	Categories
hospitalization	All	-
patient	All	-
adt	All	location_category
labs	lab_result_dttm, lab_category, lab_value	albumin, creatinine, lactate, etc.
vitals	recorded_dttm, vital_category, vital_value	heart_rate, map, sbp, spo2, temp_c
patient_assessments	All	gcs_total
medication_admin_continuous	med_dose, med_dose_unit	vasopressors
respiratory_support	All	device_category, fio2_set, peep_set

---

Configuration

Create clif_config.json from template:

{
    "site": "your_site_name",
    "data_directory": "/path/to/clif/data",
    "filetype": "parquet",
    "timezone": "US/Central"
}

---

Metrics

Classification (Tasks 6, 7)

• AUROC, AUPRC, F1, Accuracy
• Precision, Recall, Specificity, NPV
• Brier Score, ICI (Integrated Calibration Index)
• DCA (Decision Curve Analysis)
• Bootstrap 95% confidence intervals

Regression (Task 5)

• MSE, RMSE, MAE, R²
• Predicted vs Observed plots

---

Dependencies

• Python 3.11+
• UV package manager
• clifpy >= 0.3.4
• marimo notebooks
• XGBoost, scikit-learn, pandas, polars

---

Support

For issues or questions: - Check RUN.md troubleshooting section - Contact the FLAME-ICU team via CLIF consortium channels