EquiEmbryo AI

Inspiration

AI models assist IVF clinics with embryo selection, but real clinical data contains structural inequalities. We were inspired to explore how biased training labels can lead to AI systems that unintentionally disadvantage certain demographic groups in reproductive healthcare.

What it does

Our project builds a fairness auditing tool for IVF embryo-ranking AI. It takes the model’s predictions and computes certain fairness metrics, such as a True Positive Rate (TPR), False Positive Rate (FPR), and accuracy to determine fairness gaps across demographic groups. This makes it easy to see when an AI model is treating one group differently from another and whether retraining actually reduces bias.

How we built it

• Python, NumPy, Pandas, scikit-learn • Created fairness evaluation methods • Validation with Fairlearn • Visualizations using Matplotlib • Built and tested inside VS Code + Jupyter

Challenges we ran into

• Making a dataset that realistically showed how bias can happen. • Getting the fairness calculations to work the way we expected. • Realizing that a model can look accurate overall while still being unfair.

Accomplishments that we're proud of

• Building a tool that clearly shows when a model is treating groups differently.
• Fixing the bias and proving it with simple graphs.
• Understanding the full process from creating data → training models → checking fairness.

What we learned

• Even small mistakes in training data can cause big differences in how groups are treated.
• Improving fairness improve accuracy.
• TPR, FPR, and accuracy rates can determine disparities between AI model predictions.

What's next for EquiEmbryo AI

• Adding more fairness checks so it can catch different types of bias.
• Testing the tool on real embryo ranking models.
• Making a simple web demo where anyone can upload predictions and see the fairness results.