Improving Heart Disease Probability Prediction Sensitivity with a Grow Network Model

Abstract

The traditional approaches in heart disease prediction across a vast amount of data encountered a huge amount of class imbalances. Applying the conventional approaches that are available to resolve the class imbalances provides a low recall for the minority class or results in imbalance outcomes. A lightweight GrowNet-based architecture has been proposed that can obtain higher recall for the minority class using the Behavioral Risk Factor Surveillance System (BRFSS) 2022 dataset. A Synthetic Refinement Pipeline using Adaptive-TomekLinks has been employed to resolve the class imbalances. The proposed model has been tested in different versions of BRFSS datasets including BRFSS 2022, BRFSS 2021, and BRFSS 2020. The proposed model has obtained the highest specificity and sensitivity of 0.74 and 0.81 respectively across the BRFSS 2022 dataset. The proposed approach achieved an Area Under the Curve (AUC) of 0.8709. Additionally, applying explainable AI (XAI) to the proposed model has revealed the impacts of transitioning from smoking to e-cigarettes and chewing tobacco on heart disease.