In a large-scale analysis, more than half of transplant recipients developed chronic graft-versus-host disease and 15% died from causes other than cancer recurrence. These numbers capture an unsettling truth about allogeneic hematopoietic cell transplantation. Treatment can have a long, uncertain tail.
Chronic graft-versus-host disease (cGVHD) occurs when the donor’s immune cells attack the patient’s healthy tissue. It can harm your skin, eyes, mouth, joints, and lungs. Moreover, it remains a major cause of debilitating disease and non-relapse mortality after transplantation.
A research team led by Sophie Paczesny, MD, of the MUSC Hollings Cancer Center, collaborated with Dr. Michael Mertens and Dr. Brent Logan of the International Blood and Marrow Transplant Research Center at the Medical College of Wisconsin. The team set out to detect the danger early.
Their new study, published in the Journal of Clinical Investigation, describes BIOPREVENT, a machine learning tool designed to estimate a patient’s future risk of developing cGVHD and dying without recurrence. They used blood biomarkers and routine clinical factors.
A quiet few months after being discharged from the hospital
cGVHD is usually diagnosed when symptoms surface throughout the body. However, the study claims that tissue destruction begins much earlier, immediately after donor cells enter the body. This creates a window of time in which patients feel better while their biology begins to change in the background.
“By the time chronic GVHD is diagnosed, the disease has often progressed over several months and has been silently damaging the body,” Pacesny says. “We wanted to know if we could detect red flags early enough for clinicians to intervene before patients become unwell and before the damage is irreversible.”
The researchers focused on the period 90 to 100 days post-transplant, a common checkpoint for follow-up care. Their idea was conceptually simple. If the blood is already transmitting the signal, clinicians may be able to spot the risk before the first joint stiffness, dry eyes, or skin changes force a diagnosis.
7 proteins, 9 clinical details
The research team analyzed 1,310 transplant recipients drawn from four multicenter cohorts: BMTCTN 0201 and 1202, the Dana-Farber Cancer Institute cohort, and the combined pediatric and adult cohort of trial NCT02194439. All participants had blood samples available at day 90 or 100.
In this group, 52% developed some form of cGVHD, 37% developed moderate to severe cGVHD, and 15% experienced non-relapse mortality.
The researchers measured seven previously validated plasma proteins from each sample at day 90/100. Some, including CXCL9, MMP3, and DKK3, were already associated with future cGVHD risk in earlier studies led by Paczesny. The paper also highlights IL1RL1, previously called ST2. This is because previous studies have linked a 1 ‘logn’ increase in IL1RL1 to a 1.6-2.0 times higher risk of non-relapse mortality.
They combined these biomarkers with nine clinical variables. This paper highlights the characteristics of transplantation, such as graft source. It also focuses on details such as age, underlying disease, GVHD prevention, and previous acute GVHD. Researchers discuss the value of standardized registry information submitted to the International Center for Blood and Marrow Transplant Research.
They then asked a straightforward question: Can machine learning make these early signals more useful than standard approaches?
When is machine learning useful and when is it not?
The researchers compared a wide range of techniques, from penalized Cox regression to tree-based approaches to deep learning models. They split the dataset into a 75% training set and a 25% validation set. Time-varying area under the curve measurements from 180 to 540 days after transplantation were used to determine performance.
Across the validation set, models combining biomarkers and clinical data outperformed models using clinical factors alone in predicting cGVHD. In particular, the improvement in non-relapse mortality appeared to be significant.
One model type remained near the top at all times. It is Bayesian Additive Regression Tree (BART). The team used BART as the basis for the BIOPREVENT model.
Deep learning did not have an advantage here. The authors state that the performance of the neural network approach is comparable to or worse than other methods. They suggest that the sample size may have been too small for deep learning to learn reliably.
The paper also shows that different outcomes depend on different predictors. For cGVHD risk, variables of high importance in the BART model include primary disease, transplant source, GVHD prophylaxis, age, conditioning intensity, female-male sex discordance, and biomarkers such as MMP3 and CXCL9. In non-relapse mortality, IL1RL1 rose to the top along with age, primary disease, GVHD prophylaxis, HLA matching, previous acute GVHD, conditioning intensity, and sCD163.
Turn predictions into groups that clinicians can use
BIOPREVENT aims to classify patients into high- and low-risk categories, rather than just spouting abstract probabilities.
For cGVHD, the authors report an “optimal” 360-day cutpoint for predicted cumulative incidence of 0.45. Above that threshold, the 360-day cumulative incidence of cGVHD was 58.0%, compared with 32.7% below it.
For non-relapse mortality, the reported cut point was 0.08. The 360-day non-relapse mortality rate for patients with at least the predicted incidence was 13.9%, compared to 3.5% in the low-risk group.
The team also measured practical classification tradeoffs. At day 360, the sensitivity and specificity of the cGVHD cutpoint were 64%, positive predictive value was 58%, and negative predictive value was 56%, respectively. For non-relapse mortality, sensitivity was 70%, specificity was 69%, positive predictive value was 14%, and negative predictive value was 81%.
This paper points out important issues. Even with a well-calibrated model, forcing a single threshold can misclassify certain subgroups. In their analysis, bone marrow transplant recipients had a higher false-negative rate than peripheral blood transplant recipients. The authors link this to lower baseline cGVHD incidence and fewer cases exceeding the cut point.
Practical implications of the research
BIOPREVENT is now available as a public web application. Clinicians can enter nine clinical variables and seven biomarker values starting at day 90/100 and receive personalized estimates of cGVHD and non-relapse mortality up to day 540, approximately 18 months post-transplant. This study frames it as a research and risk assessment support tool rather than a treatment guide.
The authors argue that earlier risk estimation could help target closer monitoring and shape future trials of pre-emptive approaches. They also discuss how enriching trials with high-risk patients could reduce the required sample size, as event rates increase in that subgroup.
At the same time, this study also reveals important limitations for practical use. The cohort provided only one biomarker time point as subsequent samples were not available. This model included only seven validated plasma biomarkers. The authors note that adding cellular measurements, such as specific T-cell subsets, may improve accuracy, but these measurements often have low throughput and limited biobank availability.
Because the dataset was retrospective and anonymized, we were unable to include some potentially important clinical variables, such as late acute GVHD and immunosuppressive status at the time of sampling. Finally, this study included only 27 patients who received post-transplant cyclophosphamide as GVHD prophylaxis. Therefore, the authors state that the interpretation of that subgroup remains uncertain and requires validation in a larger sample.
Despite these caveats, the study makes a clear bet. If the field learns to listen, the biology of cGVHD and transplant-related death may become decipherable sooner than clinical practice typically admits.
