Heart Disease Prediction Using Ensemble Solutions and Exploring the Effect of Target Engineering with Pearson Correlation Based Feature Selection

Heart disease stands as one of the most intricate health conditions, affecting a substantial population across the globe. The aim of this research is to add to the expanding knowledge base in the early detection of heart disease. The well-known heart disease dataset sourced from UCI Machine Learning Repository was utilized in this study. The dataset encompasses 15 distinct attributes, not counting the target variable, collected from 920 different patients. Pearson correlation was employed as feature selection approach with collinearity threshold of 85%. Binning method was used as target engineering strategy. Before target engineering, various well-known classification algorithms were employed, such as Random forest, Extra gradient,CatBoost, Bagging, Logistic regression, LightGBM, SVM, Ada boost, XGBoost, Decision tree, Gradient boosting, Naive bayes, and KNN. After target engineering another diverse set of classifier models deployed, encompassing Random forest, KNN, Logistic regression, SVM, Naive bayes, XGBoost, and Decision tree. Both soft and hard voting ensemble was employed on the best selected five models, according to accuracy metric. An artificial neural network model was also constructed and applied after the target engineering method. The ANN model was trained over 300 epochs as after that point the loss curves exhibited a nearly flat trend. At that amount of epochs, the ANN model achieved accuracy score of 99.59%, which was highest among of all the suggested model. The soft voting ensemble on top five models was able to gain accuracy, precision, recall, f1-score, MAE, MSE, RMSE, RAE, and RRSE score at 90.12%, 90.08%, 90.12%, 90.08%, 9.88%, 9.88%, 34.32%, 28.40%, and 77.34% respectively, which was the second highest.