Multi-Objective Feature Selection Using Non-dominated Sorting Mechanisms and Bi-Directional Elimination for Heart Disease Classification

Abstract

Heart Disease (HD) is the primary cause of death worldwide, and its early detection poses challenges for physicians due to a high probability of false positives resulting from the large number of features. To address this issue, Feature Selection (FS) is being employed to identify the most significant feature subsets for HD classification. FS aims to accomplish two main objectives: (1) reducing the number of features and (2) enhancing the classification performance of the selected features. However, most current approaches treat FS as a single objective, combining these aims into one. Consequently, this research introduces a multi-objective FS approach that treats the number of selected features and classification performance as distinct objectives to avoid conflicts arising from a single objective. The study utilizes the Non-dominated Sorting Genetic Algorithm (NSGA-III), a bi-directional elimination-based method, as the optimization algorithm during the search process. The proposed method is evaluated on nine HD datasets of varying complexity, employing five Machine Learning (ML) classifiers as evaluation measures. The results demonstrate that the proposed method performs favorably on most datasets in terms of both the number of selected features and classification performance. The accuracy of each ML classifier surpasses its respective Sensitivity, Specificity, and F-Measure. Support Vector Machine (SVM) excels in terms of the number of selected features, while Decision Tree Classifier (DTC) exhibits superior classification accuracy. K-Nearest Neighbors (KNN) yields promising results for high-dimensional datasets. Comparative analysis with existing studies establishes the superiority of the proposed method.