Model Explaninability (SHAP and Feature Importance)

[joalen]

Building an ML model (or more specifically a classifier for Heartcheck), while fantastic for giving the result, doesn't account for what caused such a result to happen. Say, we tell a patient they have heart disease, but they don't know what caused it. Perhaps it can be resolved or mitigated further to lessen the chances of suffering further complications for heart disease.

1. Here, we'll use SHAP (SHapley Additive exPlanations) to quantify how much each feature "input" contributed to a particular prediction
2. In my probability and statistics class, I'm currently learning about confidence intervals, so I'll use that to make the model more quantifiable (things like, "No, you don't have heart disease with a 95% confidence interval between 15-25%)
3. Counterfactuals -- allow patients to see if certain inputs (or variables) change and see if that changes the prediction (like a risk would drop from 20% to 10% provided if we changed heart rate or blood pressure)

This gives more trust and accuracy in the model