Figure 3.

Row hierarchies. Internal nodes of the hierarchy can be collapsed resulting in consensus rows (which are assigned unique numerical labels starting with a capital 'A'). For nominal values, the relative frequency of the character in the consensus is mapped to saturation of the respective color. For ratio-scaled values, the mean is used instead. The row-hierarchy has been created using the automatic aggregation feature: (A) internal nodes at depth one correspond to rows with the same symptoms with children grouped by mortality. (B) Collapsing internal nodes at level 2 show the consensus of rows with the same value for symptoms. Hiding labels improves the visual pattern matching due to uncluttering, as we can discover columns with the same trend of saturation.

Heinrich et al. BMC Bioinformatics 2012 13(Suppl 8):S2   doi:10.1186/1471-2105-13-S8-S2