The Bias Destroyers

Inspiration

We were inspired by the large problem space of healthcare AI. From our observations, there were very few available resources regarding how SDOH and protected class information effects healthcare outcomes. Given this, we figured that healthcare models may not be trained as efficiently as they could be, which inspired our project.

What it does

The code submission takes a CSV file of protected class data and their health outcomes and runs tests on it using IBM's bias360 model. This helped us determine the proper outcomes.

How we built it

We built this by identifying datasets which contained protected classes that we could debias with. From there, we trained random forest models, XGBoost algorithms, and logistic regression models to make predictions on the data. From there, we used AI Fairness metrics to identify bias within the predictions with respect to each protected class. We focus on group fairness prediction methods for this. From there, we use debiasing methods such as reweighting in order to create a final debased machine learning model.

Challenges we ran into

We ran into many technical challenges. Our team did not have a strong web developer, so both of us tried to learn web development on the fly while working on debiasing models. This proved to be very challenging and ultimately was a large setback for our work.

Accomplishments that we're proud of

The toolkit that we built works to identify bias with one dataset, which we see as a good plus.

What we learned

We learned quite a bit about project planning. As this was our first hackathon, we had many ambitions for what we hoped to accomplish. However, we quickly learned that it takes a lot of time to go through the data and prepare properly.

What's next for The Bias Destroyers

We are very passionate about the issue and will continue research about AI fairness in healthcare.

Built With

• aif360
• python
• scikit-learn