EBPSO: Enhanced binary particle swarm optimization for cardiac disease classification with feature selection

Abstract

AbstractCardiac disease is one of the leading causes of death worldwide, and its early detection and diagnosis can considerably increase the lifespan of patients. An automated expert system for early and accurate diagnosis that offsets human error can be designed using machine intelligence and appropriate pre‐processing of data to ensure accuracy. To that end, enhanced binary particle swarm optimization (EBPSO) has been investigated in this paper to enable the definitive classification of cardiac disease with the aid of feature selection. The primary objective of this study is to improve the accuracy of the classification and the convergence speed while employing a lesser number of features. In order to achieve this, a new adaptive inertia weight was introduced to balance the exploitation and exploration capability of binary particle swarm optimization (BPSO). In addition, a novel idea called ‘neighbourhood best’ is introduced in the velocity update equation to improve the convergence speed. The proposed EBPSO approach was tested over well‐known datasets of heart disease, that is, the Cleveland dataset, Hungarian dataset, Switzerland dataset, and Long‐Beach‐Va datasets. For classification purposes, two well‐known classifiers, k‐nearest neighbour, and support vector machine were used with the EBPSO. Furthermore, the performance of EBPSO was compared with the traditional BPSO, as well as with BPSO with proposed inertia weight (Wa‐BPSO) and other PSO variants. It has been observed that the proposed approach EBPSO with KNN shows remarkable efficiency in terms of average classification accuracy and convergence rate. The average classification accuracy for Cleveland dataset is 92.814%, for the Hungarian dataset is 91.97%, for the Switzerland dataset is 91.626% with EBPSO+KNN and for the Long‐Beach‐Va dataset is 85.253% with EBPSO + SVM. Additionally, the performance of Wa‐BPSO was compared with other commonly used inertia weights. The results of this study demonstrated that the proposed approach produces better results in terms of both the accuracy of classification as well as the convergence rate with a reduced feature set.