Prediction of Heart Disease Using Random Forest and Rough Set Based Feature Selection

1. Introduction

Data mining plays a significant role in extracting the hidden patterns from datasets of medical domain and these patterns helpful for early detection of heart disease, which is the cause of approximately 12 million deaths occur across worldwide. Coronary heart disease deaths are observed more in the United States than in any other developed and developing countries (Soni, 2011). Coronary heart disease is also known as coronary artery disease (CAD) due to the fact that plaque or a wax like substance gets accumulated and contracts the coronary arteries. These arteries supply blood and oxygen to heart. This, in turn, affects the functioning of the heart and other organs. Atherosclerosis is one of the ailments which occur when a substance like a plaque or fatty material builds up on the artery walls. Though women in their 40s have a lower risk of CAD but when compared to men, their risk factor increases as they grow in age. Some of the diseases/attributes that increase the risk factor in women are but not limited to a) high LDL b) high BP c) diabetes d) smoking e) cholesterol f) obesity (Jabbar, 2015) and along with the major risk factors are Electrocardiographic pattern of left ventricular hypertrophy, Elevated serum cholesterol, and hyper-tension. These factors signify that there is a chance of coronary heart disease (CHD) (Kannel, 2011). CHD is a major cause to increased mortality rate across worldwide (Chaurasia, 2013).

Data mining plays a vital role in health sector for prediction of heart disease (Vijiyarani, 2013). Rough Set Theory (RST) was introduced in 1982 as a methodology for data analysis. RST works on the principle of in-discernibility relation i.e. the ability to find the difference between objects, based on their attribute values. Here, we construct upper approximations (UP) and lower approximations (LU). Lower approximation (LU) represents the objects that surely belong to a given response class, whereas upper approximation is those objects that possibly belong to the decision class. An important factor of RST is that it doesn’t require any additional parameter for analysis of data RST has been used for feature selection, instance selection, classification and regression (Riza, 2014). Feature Selection (FS) is a pre-processing step in machine learning; it helps to categorize the different categories such as relevant, irrelevant and redundant which indirectly increases learning accuracy (Jabbar, 2013). Once the data is categorized, redundant and irrelevant features are removed. Feature selection (FS) is one of the crucial steps in the medical industry for prediction of medical diseases (Jabbar, 2013).

Random Forests is the most popular ensemble technique used for prediction and probability estimation. In Random Forests N-number of decision trees are generated and each decision tree is selected based on a vote, then the decision of the class is considered. It is an excellent method for handling huge amounts of data and missing values (Jabbar, 2016). The paper is organized as follows – Section 2 discusses the earlier work done by researchers and our work in this perspective. Section 3 describes different Machine learning concepts, ensemble technique –Random Forests, Feature selection and Rough Set. Section 4, represents the proposed method for early diagnosis of heart disease using Rough Set and Random Forests. Experimental results are discussed in section 5 and concluding remarks in section 6.