A Visualization Methodology for Studying Relations of Medical Data via Extended Dependency Networks

Simon Fong, Luke Lu, Jinan Fiaidhi, and Sabah Mohammed


Medical data visualization, Dependency Networks, Correlation, Feature Selection, RIPPER


Medical professionals are keen to investigate the relations between symptoms and diseases as well as related drugs, therapies and genes. A generic visualization methodology is proposed in this paper that covers three main tools for studying the relations between attributes and predicted outcomes. They are namely Network Graph which visualizes the strengths of the links (intra-relations) between each pair of attributes within a single disease; Dependency Network that lays out all the attributes and their respective predictive powers to a disease(s), also inter-relations between symptoms across different diseases can be inferred; a rule-based Decision Tree is used to predict an outcome of a disease given an new instance of attributes. Network Graph and Decision Tree have been studied individually in the past as standalone tools. Our main contribution, despite the unifying approach for combining the three applications, is the ensemble feature selection analysis that technically enables constructing compact and accurate decision tree. The same output from the feature selection process is used to fuel building a dependency network by assigning the attributes of a diseases significance values. Furthermore we extended the dependency network from a single predicted class to multiple, which allows indirect relations between attributes across a chain of related diseases to be formulated.

Important Links:

Go Back