Page 203 - 2019_01-Haematologica-web

Machine learning refines cGvHD classification
A
Figure 3. A simple, physician- driven decision tree defines chronic graft-versus-host dis- ease phenotypes. A decision tree designed to separate patients into groups with sim- ilar phenotypes and clinical risks as those revealed by the machine-learning approach in Figure 1 is shown. The decision tree is read from the top down and sequentially identifies and segregates patients in the most pheno- typically distinct clusters (Y=Yes, N=No). Patients meeting the criteria at the decision point are assigned to that cluster and patients who do not meet the criteria are further advanced in the tree logic. Each circled num- ber represents a cluster of patients. For cluster 2, two decision points were used to identify patients (arrows above and below the encir- cled 2). The length of the hor- izontal arrow is proportional to the risk coefficient and the width of the arrow is propor- tional to the percentage of patients in this cohort who were assigned to the cluster.
Figure 4. A simple, physician-dri-
ven decision tree created groups of B patients with chronic graft-versus- host disease that were similar to computational patient clusters and stratified for overall survival. (A) Cluster numbers, newly calculated marker enrichment modeling (MEM) labels, phenotype interpre- tations (italics), risk coefficients, and group frequencies (n=339) are shown for the new groups of patients defined using the decision tree in Figure 3. MEM labels and risk were calculated as before (Figure 1 and Methods). Phenotype interpretations were assigned by expert physicians based on analy- sis of MEM labels and risk. Decision tree groups 1-3 were lower risk, groups 4-5 were intermediate risk, and groups 6-7 were higher risk. (B) Overall survival probability was stratified for patients with chronic graft-versus-host disease identified in the low-, intermediate-, and high-risk groups defined by the
physician-driven decision tree.
haematologica | 2019; 104(1)
193