A Study of Dimensionality Reduction’s Influence on Heart Disease Prediction

Abstract

Heart disease is a serious threat to human life due to its suddenness and ponderance. It’s urgent and meaningful to build a diagnosis system to detect heart disease earlier and accurately. In the field of medicine, doctors have summarized lots of experience on heart disease diagnosis. Duo to a large number of samples and attributes, the work done by the human is not efficient. And, computer-aided disease diagnosis has shown its advantages. Many researchers have applied machine learning methods to heart disease detection. For pursuing better performance, dimensionality reduction methods are often used for selecting key features or accelerating the processing speed. In this research, we investigate the influence of dimensionality reduction by using PCA and LDA methods on the machine learning methods’ prediction. PCA and LDA represent two famous dimensionality reduction, unsupervised and supervised methods. The results display that the performance of PCA is better than LDA’s evaluated by several metrics. Additionally, PCA indeed promotes many different methods’ prediction effects. There is an optimal amount of features when using PCA. It seems that the dataset with more features is easy to obtain better results. Otherwise, the dataset itself also has a significant influence on prediction result even their structures are the same. The dimensionality reduction will influence the time consumption of machine learning methods. Finally, we reveal that complex models are not always better than simple ones.