Automatic Heart Disease Prediction Using Feature Selection and Data Mining Technique

This paper presents an automatic Heart Disease (HD) prediction method based on feature selection and data mining techniques using provided symptoms and clinical information in the patient’s dataset. Data mining which allows the extraction of hidden knowledges from the data and explores the relationship between attributes, is the promising technique for HD prediction. HD symptoms can be effectively learned by the computer to classify HD into different classes. However, the informationprovided may include redundant and interrelated symptoms. The use of such information may degrade the classification performance. Feature selection is an effective way to remove such noisy informationmeanwhile improving the learning accuracy and facilitating a better understanding for learning model. In our method, HD attributes are re-selected based on their rank and weights assigned by Infinite LatentFeature Selection (ILFS) method. Support Vector Machine (SVM) algorithm is applied to classify a subset of the selected attributes into different HD classes. SMOTE (Synthetic Minority Over-sampling Technique) data over-sampling technique is adopted to generate more amounts and varieties of data. The experiment is performed on the UCI Machine Learning Repository Heart Disease public dataset. Experimental results demonstrated that by only using a subset of selected 24 attributes over a total of 46 attributes, our method achieved an accuracy of 97.87% for distinguishing ‘no presence’ HD with ‘presence’ HD and an accuracy of 93.92% for distinguishing 5 different classes of HD.