Impact of categorical and numerical features in ensemble machine learning frameworks for heart disease prediction

Cardiovascular disease (CVD) or heart disease is one of the most fatal diseases of the world that has been observed through-out the last decade. The prediction of CVD in majority of cases depends on a set of combination of clinical and pathological data represented by either numerical or categorical variables. The categorical medical data may inherently embed prior medical information during its process of categorisation. Whereas the numerical data are flexible for accurate measurements and reading. Hence it is necessary to asses the impact of categorical and numerical features for CVD prediction. In this work, an exhaustive analysis of numerical, categorical and combination of both types of features have been done in context of state-of-the-art machine learning algorithms. The work has compared the boosting algorithms such as Gradient Boosting, Extreme Gradient Boosting (XGBoost), AdaBoost, CatBoost and additionally artificial neural networks, random forest, support vector machines (SVM), decision tree and logistic regression. A soft voting ensemble mechanism with learning algorithms has also been implemented to predict CVD. The current work has used a publicly available and widely used benchmark dataset: Cleveland heart disease dataset (UCI repository). It uses ten different performance metrics which consistently demonstrate that the categorical features outperforms the numerical and combined features. It is further observed that the ensemble learning of SVM + AdaBoost classifiers with categorical features produces optimum performance of CVD prediction.