Multi-Modal Consensus Clustering to Identify Phenotypes of Kidney Transplant Donors and Recipients and Their Association With Survival

Abstract

Kidney transplantation is an essential treatment option for individuals diagnosed with End-stage renal disease (ESRD). Being able to predict the survival of the transplant and the outcome of the recipient is an important decision point at the time of kidney allocation. Understanding the underlying characteristics of donors and recipients—referred to as phenotypes—can help in matching donors and recipients to improve patient and allograft survival. In this thesis, we are studying clustering methods to identify clusters of homogeneous donors and recipients with respect to their clinical characteristics, and using the generated phenotypes to study their relationship with kidney transplant outcomes. The dataset is a combination of both categorical and numerical data, consisting of 25824 records of donor and recipient features spanning 3 years (2009 - 2011). We investigated multi-modal clustering methods to handle the mixed data types. Two base clustering methods, KAMILA and Mixture Model, were applied resulting in 3 clusters. Consensus clustering was next applied using three consensus functions, k-modes, Majority Voting and Latent Class Analysis (LCA), to generate the final consensus-driven clusters. Latent Class Analysis (LCA) gave us the best clusters on the basis of internal evaluation indices and t-SNE visualizations. Self-Organizing Maps (SOMs) with hierarchical clustering was applied to validate the consensus clustering results. The generated clusters were evaluated by domain experts for clinical utility and each of the phenotypes. Importantly the clusters showed strong and differential associations with transplant outcomes. Some non-outcome attributes were also separately distributed across clusters.