This project presents a robust machine learning framework to address bias in medical predictive modeling, utilizing the UCI Dataset and the 2019 PhysioNet Dataset. Aimed at enhancing the predictive accuracy of sepsis outcomes, it employs a multi-model approach. Initially, six stacked machine learning models—Random Forest, Stacked Model (RF + AdaBoost + MLP), XGBoost + Logistic Regression, GBM + KNN + Decision Tree, MLP, and Deep Neural Network—are trained on preprocessed data, ensuring a comprehensive analysis of vital signs, laboratory values, and demographic information. The project innovatively tackles bias by evaluating gender and age group features, comparing performance metrics like F1 Score, Accuracy, Precision, and Recall. Following the identification of biases, debiasing techniques such as Synthetic Minority Over-sampling Technique (SMOTE) and Weight Balancing are applied to refine the models. The final phase involves the development of a fair and accessible web application, designed for medical professionals to utilize in clinical settings. This application, built with React, incorporates the debiased machine learning models to assist doctors in making informed decisions, thereby promoting equitable healthcare outcomes.

Team Members

Ashwin Ram

Jinze Zhao

Mehmet Uzgoren

Ben Wang

Abra Mustafa