2026 Undergraduate Research Showcase

Transparent AI in Healthcare: Achieving Dual-Level Interpretability Using Shapley Values

Document Type

Student Presentation

Presentation Date

4-24-2026

Faculty Sponsor

Dr. Liljana Babinkostova

Abstract

The integration of machine learning in cardiology offers immense diagnostic potential, yet the "black box" nature of advanced algorithms hinders clinical adoption. Medical professionals require predictive tools that are both highly accurate and strictly interpretable. Clinicians must trust a model’s underlying medical reasoning across diverse populations, while also possessing the ability to explain individual risk assessments to patients.

In this project, we developed an XGBoost model using a multi-hospital heart disease dataset and applied SHapley Additive exPlanations (SHAP) to achieve full-spectrum clinical transparency. Globally, SHAP feature importance proved the model consistently anchored predictions on biologically sound, universal risk factors—such as asymptomatic chest pain and oldpeak—across entirely different healthcare systems. Locally, SHAP waterfall plots deconstructed individual predictions, translating complex mathematical log-odds into clear, feature-by-feature patient risk profiles.

Our findings demonstrate that advanced models need not sacrifice explainability for accuracy. By bridging complex mathematics and actionable clinical insights through dual-level SHAP interpretability, this approach ensures AI can be confidently and safely deployed in real-world healthcare settings.

This document is currently not available here.

Share

COinS