Page 200 - 2019_01-Haematologica-web

J.S. Gandelman et al.
Introduction
Stem cell transplantation is an important treatment for hematologic malignancies offering a potential cure and a treatment option for advanced disease. However, chronic graft-versus-host disease (GvHD) is a major cause of mor- bidity and mortality after a transplant.1 Chronic GvHD is a multisystem disease, however its current grading system categorizes disease compositely as mild, moderate or severe.2-4 The current grading system may overlook clini- cally relevant patterns of chronic GvHD organ scores. For example, a patient with severe skin sclerosis and a patient with highly elevated liver enzymes are both classified as having severe chronic GvHD, despite starkly different clinical manifestations of the disease.3
To date, it has not been straightforward to align the National Institutes of Health (NIH) overall severity classi- fication system and biomarkers.5 There have been some associations between the severity of chronic GvHD, as determined by the NIH classification system (NIH- Severity) and biomarkers, but biomarkers have not been able to predict clinical outcomes as strongly in chronic GvHD as in acute GvHD.6-9 Previous analyses examined disease severity in individual organs and overall disease severity but have not combined organs for phenotypic clinical subgrouping.10 A phenotypic approach to classifi- cation has the potential to characterize the pathogenesis of chronic GvHD better. Furthermore, a computational workflow capable of analyzing patterns of chronic GvHD may also have the power to elucidate patterns in other dis- eases in oncology and throughout clinical medicine.
Machine learning and clustering techniques have suc- cessfully exposed patterns in medicine, including identify- ing breast cancer metastases and genetically targeted ther- apy for acute myeloid leukemia.11-15 Machine learning has the potential to find patterns in clinical data that may be missed by the human observer and traditional approaches alone.16 A potential advantage of machine learning approaches compared to traditional statistical approaches is that results can go beyond a preformed hypothesis
allowing for discovery of novel associations and clusters.17 Additionally, with high-dimensional data, such as the types and grades of organ involvement in chronic GvHD, the multiple comparisons required in conventional statis- tics can lead to false-positives, whereas a machine learn- ing-inspired approach allows for processing of multi- dimensional data.15,18,19 Furthermore, an algorithmic approach has outperformed traditional statistics in recent clinical studies.15,20
We used a computational approach to classify patients with chronic GvHD according to organ scores, identify phenotypic subgroups and stratify survival. We hypothe- sized that machine learning methods could identify dis- tinct clusters of clinical phenotypes and survival patterns among patients with chronic GvHD.
Methods
Study population and chronic graft-versus-host disease assessment
Research was conducted with informed consent, Institutional Review Board approval and in accordance with the Declaration of Helsinki. The clinical data used were from 339 patients with inci- dent chronic GvHD enrolled in the Chronic GvHD Consortium study, a pre-existing multicenter prospective observational clinical database.21 Incident disease was defined as new chronic GvHD within the 3 months preceding the first study visit and only adult patients (≥18 years of age) were included. The original cohort size was 341; three patients were excluded because of missing organ scores, leaving 339 patients in the final analysis.
Demographics and the patients’ characteristics were collected at enrollment and through abstraction from clinical charts (Online Supplementary Table S1). At enrollment, NIH 2005 consensus crite- ria scores from 0 (no involvement) to 3 (severely affected) were recorded for eye, liver, joint, mouth, gastrointestinal tract and lung. Symptom-based lung scores were used in the initial analysis. The percentage of the body surface area with erythema (% erythema) was measured. Skin sclerosis and fascia were assessed using Hopkins scores.22
Figure 1. A machine-learning workflow reveals clusters of patients with chronic graft-versus-host disease with shared organ involvement phenotypes. t-SNE/viSNE plots show organ scores (heat) for each patient (represented by a dot) on a scale where heat indicates organ involvement. Patients who are closer together are more similar while those who are farther apart are generally more different from each other. All organ domains shown were used to generate the viSNE plots, except National Institutes of Health-Severity which was not used as a parameter to generate the viSNE maps. FlowSOM clustering is shown (right) for the seven clusters of patients, with each cluster color overlaid as a dimension on the viSNE plot. For example, Cluster 7 is pink.
190
haematologica | 2019; 104(1)