Multi-view Clustering of Heterogeneous Health Data: Application to Systemic Sclerosis

Abstract

Electronic health records (EHRs) involve heterogeneous data types such as binary, numeric and categorical attributes. As traditional clustering approaches require the definition of a single proximity measure, different data types are typically transformed into a common format or amalgamated through a single distance function. Unfortunately, this early transformation step largely pre-determines the cluster analysis results and can cause information loss, as the relative importance of different attributes is not considered. This exploratory work aims to avoid this premature integration of attribute types prior to cluster analysis through a multi-objective evolutionary algorithm called MVMC. This approach allows multiple data types to be integrated into the clustering process, explore trade-offs between them, and determine consensus clusters that are supported across these data views. We evaluate our approach in a case study focusing on systemic sclerosis (SSc), a highly heterogeneous auto-immune disease that can be considered a representative example of an EHRs data problem. Our results highlight the potential benefits of multi-view learning in an EHR context. Furthermore, this comprehensive classification integrating multiple and various data sources will help to understand better disease complications and treatment goals.