Some Methods for Longitudinal and Cross-Sectional Visualization with Further Applications in the Context of Heat Maps

Abstract

OBJECTIVES: Visualization in state sequence data has been developed extensively through an R package described in Gabadinho et al. (2011). Graphics depicting states prevalent at each cross section in time can be generated for all data as well as for covariate level sets for datasets with a large number of subjects with state transitions over time. Special longitudinal sequence sets can also be carved out using similarity measures across sequences. In our work, we believed there may be latent informative images inherent in data on changes in cancer states (degrees of response , progressions, and deaths). We obtain a longitudinal as well as a cross-sectional informative image through a novel heuristic grounded in the framework of hierarchical clustering . METHODS: We used iconic known images, stripped them of all ordering information, and attempted to recover the known latent image underlying the randomly permuted data using our heuristic as well as other alternative methods such as those in Sakai et al. (2014). RESULTS: Results validate our methods. The method is demonstrated through a visual representation of changes in cancer states for two induction therapies in a cancer trial. A further application to a two-way ordering of gene sample heat maps are also presented. CONCLUSIONS: When cancer state transition graphics for competing therapies are juxtaposed, there can be a quick read of early versus late response to therapy, the depth and duration of response as well as a rough gauge of events such as progression and death. This is a good complement to quantitative inferences. For gene expression data, we hope that our methods will bring out finer distinctions in addition to presenting gross patterns in the data like those seen using prevalent methods.

Appendix: R* Code to Generate Ordered Sequences of Rows and Columns Using the Edge Clustering Method

*R Core Team 2013.