Effectiveness of Data Driven Diagnosis of Heart Disease

Early detection of heart disease can help in preventing disease progression. Several studies consider machine learning algorithms for predicting heart diseases. However, the performance of the algorithms varies depending on the datasets. This paper generates a hybrid dataset by combining two existing datasets of heart disease. The hybrid dataset takes into account the six common attributes of the individual datasets one available at UCI ML repository and the other at Kaggle. Normalization is performed to make the samples of a common attribute to be uniform. Two approaches are used to evaluate the effectiveness of the hybrid dataset. In Method A, the samples of the hybrid dataset are randomly separated into training and testing portion using holdout or cross-validation approach. In Method B, one dataset is used as training the model and the other dataset for testing the model. The parameters of different ML algorithms are optimized while applying to the hybrid dataset. Results show that for the case of Method A using holdout approach, the best classification accuracy of 73.87% is obtained by majority voting classifier with logistic regression, SVM and naïve Bayes. For the case of Method B, the best accuracy of 58.41% is obtained by the same classifier. Finally, it is shown that for the hybrid dataset, the classification accuracy is lower than that of the individual datasets.