Page 205 - 2019_01-Haematologica-web

Machine learning refines cGvHD classification
AB
Figure 6. The physician-driven decision tree recapitulates the machine-learning workflow and finds clusters with stable risk. (A) A scatter plot shows the same patients in groups resulting from the decision tree (y-axis) or computational analysis (x-axis). Patients within or touching the black boxes were those with the same group classification in both workflows (86% of patients, n=339). (B) Bootstrapping analysis revealed stability of cluster risk across ten decision-tree analysis runs using 130 of 339 randomly sampled patients. The coefficient of risk was calculated for each run of the analysis for each cluster. The standard deviation of the ten coefficients of risks was calculated and was <0.7 for all clusters, except Cluster 3.
tree. Importantly, the decision tree stratified risk of mor- tality independently of previously defined risk factors for chronic GvHD, including NIH-Severity. Notably, platelet count was a risk factor that continued to stratify risk sig- nificantly. Overall, the decision tree has the potential to be applied in the clinical setting to assess patients’ pheno- types, once further validation in prospective, independent cohorts has been completed. Additionally, this decision tree can be applied in the research setting to large cohorts of patients.
Disease trajectory differed in the decision-tree-identi- fied clusters, most notably for Clusters 2, 6 and 7. The time from stem cell transplantation to development of chronic GvHD was different in Cluster 2, a sclerotic phe- notype. This is a clinically relevant and potentially biolog- ically distinct cluster of patients. Longer time to chronic GvHD development is a known clinical finding in patients with sclerotic chronic GvHD.5,28 Previous work defined patients with sclerotic chronic GvHD as having at least one of the following: sclerosis, fascia or joint involvement.29,30 This literature did not comment on the sclerotic phenotype as one with “de-enrichment” of liver and mouth involvement or take into account the combina- tion of multiple sclerotic features.29,30 The combination of enriched and de-enriched features we describe may enable better association with biomarkers and treatment response.
Cluster 6, a mixed phenotype, high-risk cluster, was a novel high-risk cluster revealed by the decision tree. This cluster was defined by enrichment for mouth, eye, and gastrointestinal tract involvement. Notably, this cluster required the highest number of questions on the decision tree to reach, indicating that it was poorly defined and required that other clusters were ruled out to find patients in this phenotypic group. Patients in this cluster had sig- nificantly worse overall survival when compared to all those in all other clusters combined. A caveat is that, in stability analysis of the machine-learning workflow, Cluster 6 was not highly stable, but it did recur through all repetitions of analysis (Online Supplementary Figure S5). The combination of these areas of organ involvement has
not been previously cited as a risk factor for adverse out- comes in chronic GvHD and should be further explored through cellular analyses for biomarkers and evaluated in continued validation cohorts.
Patients in Cluster 7 derived from the decision tree, a liver predominant-severe phenotype, also had a different disease trajectory when compared to patients in other clusters in that they had a significantly worse overall sur- vival than patients in all other clusters combined. This decision-tree-derived cluster is supported by previous research showing that severe elevation of liver enzymes is a known risk factor for adverse outcomes in chronic GvHD.10
Prognostication by clustering is distinct from prognosti- cation by individual organ scores alone. For example, in the machine-learning analysis, Cluster 5 lacked liver involvement and was a high-risk cluster, while high-risk Cluster 6 and Cluster 7 were specifically enriched for liver involvement. This supports the concept that this single organ score does not confer unidirectional low or high risk within the clusters. Furthermore, Liver+5 enrichment was seen in multiple low-risk clusters and one high-risk cluster. Clustering is unique in that it is not an individual organ score or characteristic but rather combinations of organ involvement and the specific absence of organ involve- ment that drive cluster formation and likely prognosis. Another example of this is that mouth enrichment was seen in both an intermediate-risk cluster (Cluster 4) and high-risk cluster (Cluster 6). Cluster 6, a high-risk cluster, comprises mouth, eye and liver enrichment; these individ- ual enrichment types appear in low-risk clusters but it is perhaps the combination that makes this a high-risk clus- ter. However, we cannot rule out that gastrointestinal tract enrichment, uniquely present in Cluster 6, is not the driv- ing force of adverse outcomes.
A limitation of the machine-learning approach is that it is not possible to add new patients to this analysis without shifting the current clusters. This was overcome by the decision-tree approach. Validation with an external cohort as well as comparison with other risk stratification tools for chronic GvHD31 should further strengthen the findings
haematologica | 2019; 104(1)
195