Forecasting Coronary Heart Disease Risk With a 2-Step Hybrid Ensemble Learning Method and Forward Feature Selection Algorithm

Abstract

Detecting cardiovascular irregularities in a timely manner is crucial for preventing any fatal risks. This research aims to devise an efficient forecasting algorithm for the timely prognosis of Coronary Heart Disease (CHD). The study includes a diverse sample of individuals from Framingham, Massachusetts, with varying demographic, clinical, and co-morbidity parameters. We aim to achieve this with a two-step ensemble Machine Learning model. Firstly, feature importance is integrated with conventional classifiers to build Feature Weighted Meta-Models with a Forward feature selection algorithm. Subsequently, the top-performing Meta-Models are combined to design the Hybrid Voting Models to predict the risk of CHD in a ten-year timeframe by minimizing the misclassification rate. The proposed models undergo vetting using multiple metrics, including F1 score, Matthew’s Correlation Coefficient (MCC), Misclassification Ratio (MCR), and Accuracy. Given the high cost associated with misclassification in the healthcare domain, these metrics are carefully considered. The resulting model demonstrated strong predictive capability for CHD risk, achieving an overall accuracy rate of 95.87%. The F1 score is calculated to be 0.91, the MCC is 0.83, and the MCR is 0.041. Notably, the model achieved these impressive results using only seven features, reducing the time complexity of the prediction. In comparison to conventional classifiers, our model achieved a 23.94% improvement in accuracy, and a 17.23% improvement over average Meta-models accuracy, highlighting its effectiveness in predicting CHD risk.

Keywords

• Accuracy
• ensemble learning
• feature weighted meta-models
• F1 score
• feature importance
• feature optimization