Partially Lazy Classification of Cardiovascular Risk via Multi-way Graph Cut Optimization

Abstract

Cardiovascular disease (CVD) is considered a leading cause of human mortality with rising trends worldwide. Therefore, early identification of seemingly healthy subjects at risk is a priority. For this purpose, we propose a novel classification algorithm that provides a sound individual risk prediction, based on a non-invasive assessment of retinal vascular function. So-called lazy classification methods offer reduced time complexity by saving model construction time and better adapting to newly available instances, when compared to well-known eager methods. Lazy methods are widely used due to their simplicity and competitive performance. However, traditional lazy approaches are more vulnerable to noise and outliers, due to their full reliance on the instances’ local neighbourhood for classification. In this work, a learning method based on Graph Cut Optimization called GCO mine is proposed, which considers both the local arrangements and the global structure of the data, resulting in improved performance relative to traditional lazy methods. We compare GCO mine coupled with genetic algorithms (hGCO mine) with established lazy and eager algorithms to predict cardiovascular risk based on Retinal Vessel Analysis (RVA) data. The highest accuracy of 99.52% is achieved by hGCO mine. The performance of GCO mine is additionally demonstrated on 12 benchmark medical datasets from the UCI repository. In 8 out of 12 datasets, GCO mine outperforms its counterparts. GCO mine is recommended for studies where new instances are expected to be acquired over time, as it saves model creation time and allows for better generalization compared to state of the art methods.