Fig. 10

Decision tree for disease progression. (a) Regression tree for the prediction of time phases after BDL based on mean cluster time courses (Fig. 9). Splitting rules are depicted at the respective branching points of the tree, with left branches correspond to ‘yes’, right branches to ‘no’ decisions. In addition to the cluster used in the decisions, also the best gene representatives from the respective cluster is shown above the decision rule. The regression tree classifies the data in six time classes 0 h, 6 h, 14 h, 24 h, 6d, 14d with information on mean time, range, number and percentage of samples falling in the respective class listed. In addition to the tree based on the mean cluster time courses (mean cluster), the best tree only using a single gene transcript from every cluster is shown (best gene). The best decision tree based on genes, histological, biochemical, and immunostaining factors (not shown) is highly similar to the depicted best gene tree, with the single change of using S100a4 instead of Col1a1 for the decision on cluster c4 and allowing GLDH as equally good alternative to Fn1 in c3. (b) Predictive performance of decision tree. The predictive performance of the regression tree was evaluated using all single factor combinations from the individual clusters (white), a random sample (N = 10000) of two factors from each cluster (gray), the best gene representative tree (red), and the mean cluster data (blue, trainings data). For all classes of the decision tree the histogram of predicted vs. experimental data are shown