callmefair 0.0.4

A comprehensive framework for automatic bias mitigation in AI systems

FairHealthGrid: A Systematic Framework for Evaluating Bias Mitigation Strategies in Healthcare Machine Learning

Paper Link* | Slides (soon)

Table of Contents:

1. Overview
2. News
3. Getting Started
4. Results
5. Citation
6. License
7. Related Resources

Overview*

The integration of machine learning (ML) into healthcare demands rigorous fairness assurance to prevent algorithmic biases from exacerbating disparities in treatment. This study introduces FairHealthGrid, a systematic framework for evaluating bias mitigation strategies across healthcare ML models. Our framework combines grid search with a composite fairness score, incorporating fairness metrics weighted by risk tolerances. As output, we present a trade-off map concurrently evaluating accuracy and fairness, categorizing solutions (model + bias mitigation) into the following regions: Win-Win, Good, Poor, Inverted, or Lose-Lose. We apply the framework on three different healthcare datasets. Results reveal significant variability across different healthcare applications. The framework identifies model-bias mitigation for balancing equity and accuracy, yet highlights the absence of universal solutions. By enabling systematic trade-off analysis, FairHealthGrid allows healthcare stakeholders to audit, compare, and select ethically aligned ML models for specific healthcare applications—advancing towards equitable AI in healthcare.

News

2025/1/6 Our paper got accepted on AMCIS 2025.

Getting Started*

To install and use FairHealthGrid, follow these steps:

1. Installation

You can install the package directly from PyPI:

pip install callmefair

2. Documentation

For detailed documentation and usage examples, visit: CallMeFair Documentation

Results

Key Findings

Our evaluation across three healthcare datasets reveals that preprocessing and in-processing strategies offer the best balance between fairness and accuracy, while multi-stage bias mitigation approaches risk over-correction. The framework demonstrates significant variability in bias mitigation effectiveness across different healthcare applications, highlighting the absence of universal solutions.

Main Contributions

We introduce FairHealthGrid, a comprehensive framework that addresses critical gaps in healthcare AI fairness evaluation:

1. Systematic Evaluation Framework: Exhaustive grid search optimization for model-strategy combinations in healthcare AI
2. Unified Fairness Score (𝓕): Integrates five fairness metrics with risk tolerances for comprehensive bias assessment
3. Interpretable Trade-off Maps: Context-driven model selection through fairness-accuracy visualization

Framework Benefits

• Model-agnostic design enabling evaluation across diverse ML architectures
• Context-specific evaluations prioritizing datasets with sensitive attributes
• Actionable insights for clinical deployment decisions
• Systematic methodology for selecting appropriate bias mitigation strategies

Limitations & Future Work

Current evaluations focus on single protected attributes. Future research will address intersectional bias from interacting sensitive features, expand to new data modalities, and formalize a comprehensive bias taxonomy using causal fairness frameworks.

Figure 1: Proposed approach consists of three main stages: (A) Grid Search, where multiple machine learning models (ML) and bias mitigation (BM) strategies are evaluated; (B) Computation of the Combined Fairness Score (𝓕), aggregating multiple fairness measures to quantify bias; and (C) Fairness Trade-off Mapping, which utilizes the baseline fairness score 𝓕baseline and naive mutation strategies (𝓕10–𝓕100) to generate a trade-off curve. The final trade-off map categorizes models into trade-off regions.

Citation

If you find our project useful for your research, please consider citing our paper and codebase with the following BibTeX:

@proceedings{bento2022deep,
  title={FairHealthGrid: A Systematic Framework for Evaluating Bias Mitigation Strategies in Healthcare Machine Learning},
  author={Paiva, Pedro and Dehghani, Farzaneh and Anzum, Fahim and Singhal, Mansi and Metwali, Ayah and Gavrilova, Marina and Bento, Mariana},
  year={2025},
  journal={31st Americas Conference on Information Systems, {AMCIS} 2025},
  publisher = {Association for Information Systems},
}

License

In the License section of your GitHub project or research paper, include the type of license under which your code or content is released. Briefly explain the permissions and limitations associated with that license. Common licenses include MIT, GPL, or Apache.

Related Resources

• Related Literature (ai2lab)
• Web Content