Meta-Prediction of Coronary Artery Disease Risk

Description

This repository provides the core machine learning codebase used in our Nature Medicine publication:
“Meta-prediction of coronary artery disease risk”
🔗 https://www.nature.com/articles/s41591-025-03648-0

Our study introduces a novel meta-prediction framework that integrates genetic and non-genetic factors - unmodifiable and modifiable risk profiles - into a unified prediction system for 10-year incident coronary artery disease (CAD). This repository shares key components of the ML pipeline.

📌 Cite us

Chen SF, et al. Meta-prediction of coronary artery disease risk. Nature Medicine. 2025. DOI: 10.1038/s41591-025-03648-0.

📎 BibTeX

@article{chen2025metapred,
  title={Meta-prediction of coronary artery disease risk},
  author={Chen, Shang-Fu and Lee, Sang Eun and Sadaei, Hossein Javedani and Park, Jun-Bean and Khattab, Ahmed and Chen, Jei-Fu and Henegar, Corneliu and Wineinger, Nathan E. and Muse, Evan D. and Torkamani, Ali},
  journal={Nature Medicine},
  year={2025},
  month={Apr},
  publisher={Nature Portfolio},
  doi={10.1038/s41591-025-03648-0},
  url={https://www.nature.com/articles/s41591-025-03648-0}
}

🧬 Study overview

_{^{👨‍💻 Insider trivia}}

_{^{The silhouette featured in Figure 1b isn’t just any figure — it’s based on Shaun (Chen SF), the first author and lead developer of this codebase. Fitting, since the meta-prediction includes both features about the individual… and of the individual. 😉}}

• We developed a meta-prediction framework that combines unmodifiable and modifiable factors to predict 10-year risk of CAD.
• The UK Biobank dataset was partitioned into two cohorts, each serving a distinct role in the pipeline:
  • A prevalent CAD cohort, used to train baseline models for predicting biomarker levels and diagnostic categories.
  • An incident CAD cohort, used to build the final CAD risk model based on meta-features derived from baseline model outputs.
• The framework generated 296 meta-features from ~2,000 variables, including clinical biomarkers, diagnostic categories, and >1,000 polygenic risk scores (PRSs).
• The final model used 50 selected features (13 measured variables, 22 PRSs, and 15 meta-features), achieving AUROC 0.84 in UK Biobank and AUROC 0.81 in All of Us
• The framework supports individualized intervention simulation and identifies subgroups with differential benefit, offering new opportunities for precision prevention.

🧰 Assets

This codebase includes components for training individual prediction models, such as CAD diagnoses and biomarker estimations. A consistent pipeline used across multiple prediction tasks to enable meta-feature generation, which feeds into our final CAD risk model and trained the final model.

Key features include:

• Compatible with tested tree-based ML models: XGBoost, LightGBM, CatBoost
• Custom utilities:
  • zoish: SHAP-based feature importance wrapper built on fasttreeshap
  • lohrasb: Optuna-based hyperparameter tuner (TPE + Hyperband)

⚙️ Environment Configuration

This project requires Python >=3.10 and uses Poetry for dependency management. If you prefer pip, a requirements.txt is also provided.

Option 1: Using Poetry

poetry install

Option 2: Using pip

pip install -r requirements.txt

Dependencies

Main runtime dependencies:

catboost==1.2.5
category-encoders==2.6.3
fasttreeshap==0.1.6
lightgbm==4.5.0
lohrasb==4.2.0
matplotlib==3.8.4
numpy==1.21.6
optuna-integration==3.6.0
pandas>=1.3.5
ray==2.7.1
scikit-learn==1.0.2
seaborn
shap==0.42.1
tune-sklearn==0.5.0
xgboost==1.7.5
zoish==5.0.4

🚀 Usage

You may either run the provided notebook or invoke the command-line interface:

Option 1: Run the tutorial notebook

A reproducible example is provided in Tutorial_Notebook.ipynb, demonstrating model training and SHAP analysis.

Option 2: Run the pipeline via CLI

The CLI expects a .pkl file containing a preprocessed pandas.DataFrame with appropriate data types. It must include:

• A target column matching --y_label
• An ID column matching --id_col

Sample Data

This project uses the publicly available Cardiovascular Disease dataset as an example. The input file data/typed_cardio_train.pkl is generated by data/make_cardio_train_pickle.py and tracked with Git LFS.

To explore all available options and their default values, run:

python -u meta_prediction_estimator.py -h

Example command:

python -u meta_prediction_estimator.py \
  --y_label "cardio" \
  --input_pickle_fp "data/typed_cardio_train.pkl" \
  --id_col "id" \
  --pkg "xgb" \
  --estimator_type "classifier" \
  --n_features 5 \
  --n_trials 100

The pipeline will produce:

1. A trained pipeline object (including the best estimator)
   ↳ final_pipeline__xgb_classifier__cardio.joblib

2. A SHAP-based feature importance file (mean absolute SHAP values per feature)
   ↳ shap__xgb_classifier__cardio__preselect.tsv

See expected_output/ for example outputs.