Machine learning reveals chronic graft-versus-host disease phenotypes and stratifies survival after stem cell transplant for hematologic malignancies

Abstract

The application of machine learning in medicine has been productive in multiple fields, but has not previously been applied to analyze the complexity of organ involvement by chronic graft-versus-host disease. Chronic graft-versus-host disease is classified by an overall composite score as mild, moderate or severe, which may overlook clinically relevant patterns in organ involvement. Here we applied a novel computational approach to chronic graft-versus-host disease with the goal of identifying phenotypic groups based on the subcomponents of the National Institutes of Health Consensus Criteria. Computational analysis revealed seven distinct groups of patients with contrasting clinical risks. The high-risk group had an inferior overall survival compared to the low-risk group (hazard ratio 2.24; 95% confidence interval: 1.36-3.68), an effect that was independent of graft-versus-host disease severity as measured by the National Institutes of Health criteria. To test clinical applicability, knowledge was translated into a simplified clinical prognostic decision tree. Groups identified by the decision tree also stratified outcomes and closely matched those from the original analysis. Patients in the high- and intermediate-risk decision-tree groups had significantly shorter overall survival than those in the low-risk group (hazard ratio 2.79; 95% confidence interval: 1.58-4.91 and hazard ratio 1.78; 95% confidence interval: 1.06-3.01, respectively). Machine learning and other computational analyses may better reveal biomarkers and stratify risk than the current approach based on cumulative severity. This approach could now be explored in other disease models with complex clinical phenotypes. External validation must be completed prior to clinical application. Ultimately, this approach has the potential to reveal distinct pathophysiological mechanisms that may underlie clusters. Clinicaltrials.gov identifier: NCT00637689.

Organ Changes Associated with Provider-Assessed Responses in Patients with Chronic Graft-versus-Host Disease

“…Future efforts toward this goal would benefit from methods to ensure that providers carefully compare serial values to minimize recall bias before reaching conclusions about overall improvement or worsening. Future efforts would also benefit from more complex analytic methods such as machine learning that could accommodate different weights assigned to unit changes from baseline anchors across the range of each scale [20]. The success of such efforts will certainly be limited by the inherent variability of provider judgment even when evaluating the same clinical scenario.…”

A Systematic Review of Machine Learning Techniques in Hematopoietic Stem Cell Transplantation (HSCT)

“…Lopez et al [29] used random forest to identify genes associated with the incidence of chronic GVHD while Sharifi et al [30] used unsupervised methods to distinguish pulmonary complications post-HSCT. Gandelman et al [31] also utilized unsupervised machine learning techniques for risk stratification of chronic GVHD. They first converted multidimensional transplant recipient data into two dimensions using vi stochastic neighbor embedding (viSNE), then applied a self-organizing maps (SOM) algorithm for patient clustering based on organ scores.…”

High risk glioblastoma cells revealed by machine learning and single cell signaling profiles

“…After patterns are recognized by a machine learning approach, it can be valuable to determine whether the learned features can be identified using simpler models that can be applied by experts or machines to new datasets. One approach is to create a decision tree using one-or two-dimensional gating 23 , consistent with traditional strategies in immunology and hematopathology. Such gates make the identification of cells computationally less intensive and more pragmatic for wide-spread clinical use and have been previously used in glioblastoma mass cytometry 24 .…”

“…These tools help explore the structure of multidimensional data and reveal subpopulations that can be overlooked in expert manual analysis 10,12,13,22 . However, while it is possible to quickly review enriched features of the groups 21 , it would also be powerful to test whether groups with similar phenotypes share an association with differential risk of death 23 . RAPID, a fully unsupervised workflow presented here, implements t-SNE, FlowSOM, and MEM analysis of single cell mass cytometry data to reveal risk stratifying cell populations 23 .…”