DATA MINING MODELS FOR CAD USING NON INVASIVE CLINICAL PARAMETERS

Abstract

Cardiovascular diseases (CVD) are becoming common with changing life styles across the world. It majorly happens due to various disorders of heart and blood vessels, and is considered as one of the foremost reason of death and disability. Early disease diagnosis and treatment can reduce the threat of having further severity of the disease, and hence, associated mortality. With the availability of the clinical data of probable and potential patients, it becomes imperative to find suitable models to diagnose these diseases both economically and accurately. One of the most common CVD is Coronary Artery Disease (CAD) that infests in presence of atherosclerotic plaques in coronary arteries restricting the flow of blood to the heart muscle by physically clogging the artery, leading to cardiac arrest or myocardial infarction. CAD patients can be diagnosed accurately using angiography which is an invasive, costly and highly technical procedure. Angiography can be risky, may lead to further aggravation of the disease and is not suitable for screening of larger population or for a follow-up of a subject immediately after treatment or surgery. Due to the limitations of CAD diagnostic methods, researchers are seeking other methods that are less expensive, less risky, less complex, fast and easily reproducible. This work explores various data mining techniques to model diagnosis of CAD and its severity. newlineThe benchmarked Cleveland heart disease data from UCI machine learning repository and CAD clinical data from the Department of Cardiology, Indira Gandhi Medical College, India were utilized for the experimental purposes. Data set consist of 26 features. newline newlineA hybrid data mining frame-work is developed for detection of CAD using only noninvasive clinical parameters. Risk factors are identified using correlation based feature subset selection with particle swam optimization (PSO) search method. A K-means clustering algorithm is applied before the model construction. Supervised learning algorithms such as multinomial logistic regression (MLR), multi-layer perceptron (MLP), C4.5 and fuzzy unordered rule induction (FURIA) are then used to model CAD cases. The hybridization improves accuracy of algorithms by 8.3% to 11.4% for the Cleveland data set. Furthermore, another novel cluster based method is used for missing value imputation and then, a fuzzy rule based classification newlinesystem is explored to model the severity of CAD cases. Significant diagnostic accuracy is achieved by fuzzy rule based classification system. These methods can serve as an adjunct tool to screen CAD cases without using invasive clinical diagnosis and help a medical practitioner to prescribe appropriate treatments, as the case may be. newline newline