Inspiration

Cardiovascular disease remains the leading cause of global mortality. Yet predictive healthcare models often optimize for overall accuracy while overlooking who the model performs well for. While reviewing cardiovascular datasets, we noticed something subtle but critical: a model can achieve an AUC above 0.90 and still disproportionately miss diagnoses in specific demographic groups. In clinical screening, a false negative isn’t just a statistical error, it can delay treatment, increase complications, and cost lives. That realization shaped FairHeart.

What We Learned

This project reshaped how we think about model performance. High aggregate metrics can mask subgroup disparities. Fairness in healthcare AI must prioritize error distribution, not just overall accuracy. Post-hoc mitigation techniques can meaningfully improve equity without retraining complex models. Ethical AI design requires both statistical rigor and clinical reasoning

Built With

• machine-learning
• matplotlib
• python
• scikit-learn