Vitalpath: A Cardiovascular Risk Assessment Framework Using Bayesian-Optimized Ensembles and SHAP

Abstract

Heart disease is one of the leading factors of death in whole world. Yet, predicting it early remains a huge setback. Doctors face two main problems in hospitals: patient files usually contain empty fields and complex AI models act like black boxes, making them hard to trust for users. VitalPath AI is designed to solve these issues in efficient and reliable manner. First, we tackle data gaps using the MICE algorithm. This lets us fill in missing patient details without throwing away valuable data. Next, we use SMOTE to balance the dataset classes and level the playing field, ensuring the model learns fairly and prevents any model from generating biased outcomes. Instead of just guessing settings, we used Bayesian Optimization to hunt down the optimal configurations for several machine learning models. When evaluated on the UCI Heart Disease dataset, AdaBoost came out on top by ending up attaining an AUC-ROC score of 0.963. This performance metrics surpassed what we originally hoped for, though accuracy alone isn't the only objective for this work to accomplish. To make the model easy for doctors to trust, we need transparency. That’s why we integrated SHAP, which breaks down exactly why each prediction was made, letting doctors and patients see if factors like cholesterol or chest pain drove the decision.