Augmented visualization of class–cluster match in patient similarity networks

Abstract Visual representation of data mining results is essential for accurate assessment by domain experts without extensive knowledge of complex relationships in multi-label data. Similarity networks are a commonly used tool for analyzing such data, utilizing data clustering and community detection and visualizing the relationships between individual objects in an understandable form. In clinical data mining, a patient similarity network (PSN) can help clinicians identify patient clusters with representative labels and interpret their relationships. This article demonstrates the use of the Matthews correlation coefficient (MCC) to analyze cluster-class relationships to complement the PSN visualization in several synthetic datasets. We then discuss the limitations of MCC for this application and propose a modification in the form of a rescaled MCC (rMCC). Furthermore, we introduce a novel measure, Connection Purity, that complements rMCC in an informative way. We propose an augmented visualization of patient similarity networks utilizing both measures. We demonstrate this approach on several real-world datasets, showing how clinical intuition may be biased and how our method helps to rectify it.