Warfarin Dose Prediction With Explainable ML

ISEF Category: Computational Biology and Bioinformatics

Subcategory: Computational Pharmacology  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

A warfarin dose that is right for one person can be too weak, or too strong, for another. That matters because warfarin affects blood clotting, so the wrong dose can cause bleeding or clots. Your job is to see whether a machine learning model can predict dose well, and whether it treats different ancestry groups fairly.

What Is It?

Warfarin is a medicine doctors use to prevent dangerous blood clots. The tricky part is that people need very different doses. Age, weight, sex, other drugs, and genetics can all change the right amount. Think of it like tuning a radio. One setting does not work for every station, and one dose does not fit every patient.

An ML dose-individualization model tries to learn those patterns from past patient data. You feed in features such as age, body size, genetic markers, and clinical factors, then the model predicts a dose. Explainability tools like SHAP, which stands for SHapley Additive exPlanations, help you see which features pushed a prediction up or down. That makes the model less of a black box and more of a system you can inspect for errors and bias.

Why This Is a Good Topic

This is a strong science fair topic because you can test a real medical problem with public data and no wet lab work. The question is clear, the metrics are measurable, and the analysis can go beyond simple accuracy. You can learn data cleaning, model training, feature importance, and fairness checking, all skills that matter in computational biology.

Research Questions

• How does adding pharmacogenomic features change warfarin dose prediction error compared with clinical features alone?
• What is the effect of ancestry group on model error after controlling for age, body size, and genotype?
• Does SHAP identify different feature drivers for dose prediction across ancestry groups?
• To what extent does model performance change when you train on one ancestry subgroup and test on another?
• Which model type, such as linear regression, random forest, or gradient boosting, gives the best balance of accuracy and interpretability?
• What is the effect of removing ancestry proxy variables on prediction accuracy and fairness metrics?

Basic Materials

• Laptop or desktop computer with at least 8 GB RAM.
• Internet access for downloading the IWPC public dataset and reading background papers.
• Python installed with Jupyter Notebook or Google Colab access.
• Spreadsheet software for quick inspection of the dataset.
• Free account access to GitHub, if you want to track code changes.
• Reliable note-taking system for recording feature definitions, missing values, and data splits.

Advanced Materials

• Laptop or desktop computer with at least 16 GB RAM.
• Python environment with scikit-learn, pandas, numpy, matplotlib, seaborn, and SHAP.
• Jupyter Notebook or JupyterLab for exploratory analysis.
• R with tidyverse and caret, if you want to compare workflows.
• Access to a secure folder for versioned data cleaning logs.
• Statistical testing tools for bootstrapping, confidence intervals, and subgroup comparisons.

Software & Tools

• Python: Runs data cleaning, model training, and fairness analysis for the warfarin dataset.
• Jupyter Notebook: Lets you document code, figures, and model checks in one place.
• scikit-learn: Provides baseline and tree-based models for prediction and cross-validation.
• SHAP: Estimates how each feature moves a dose prediction up or down.
• Google Colab: Gives you a free cloud notebook if your computer is slow.

Experiment Steps

1. Define the exact prediction target, your main outcome metric, and the subgroup fairness metrics you will compare.
2. Inspect the IWPC dataset, then decide which clinical, genetic, and ancestry-related features are allowed in each model version.
3. Build a baseline model first, then compare it with a more complex model so you can separate signal from overfitting.
4. Plan a train-test split that protects against data leakage and keeps subgroup comparisons honest.
5. Add SHAP analysis to rank feature influence, then check whether the explanation pattern changes across ancestry groups.
6. Design a fairness analysis that compares error distributions, not just average accuracy, across patient subgroups.

Common Pitfalls

• Mixing training and test data during preprocessing, which makes the model look better than it really is.
• Treating ancestry as a simple label without checking whether the feature acts as a proxy for other missing clinical variables.
• Reporting only overall accuracy, which can hide large dose errors inside a subgroup.
• Using SHAP values on a model that was not set up consistently, which makes the explanation unstable across runs.
• Ignoring missing genotype or clinical fields, which can shrink the dataset and bias the subgroup comparison.

What Makes This Competitive

A strong version of this project does more than predict dose. It compares multiple models, tests subgroup error patterns, and explains why the model behaves differently across ancestry groups. You can raise the level by using careful validation, uncertainty estimates, and fairness metrics, not just a single score. A very strong project asks whether explainability tools reveal a real equity problem, and then checks that finding with more than one analysis method.

Project Variations

• Use a different public anticoagulation dataset and see whether the same ancestry fairness pattern appears.
• Compare SHAP with permutation importance to see whether the explanation changes when the model changes.
• Replace ancestry labels with genetic ancestry proxies and test whether prediction error shifts in the same way.

Learn More

• PubMed: Search for review articles on warfarin pharmacogenomics, dose prediction, and model fairness in clinical ML.
• NIH Genetic and Rare Diseases Information Center: Read plain-language background on genetics and drug response.
• NIH PubMed Central: Find full-text papers on warfarin dosing models and explainability methods.
• NIH All of Us Research Program: Explore how large-scale health data studies discuss ancestry and representation.
• scikit-learn Documentation: Review model building, validation, and cross-validation methods for tabular data.
• SHAP Documentation: Learn how to compute and interpret feature attributions for machine learning models.