PF749 A MACHINE‐LEARNING APPROACH TO GRADE ACUTE GRAFT‐VERSUS‐HOST DISEASE

Abstract

Background:Acute graft‐versus‐host disease (GVHD) remains a major cause of morbidity and mortality after allogeneic hematopoietic stem cell transplantation (HSCT). After the presentation of the first GVHD grading score by Glucksberg et al. in 1974, revised by the Seattle group in the 1990 s, several studies have tried to further improve severity indexing (SI) of acute GVHD based on different evaluations of organ involvements. Today, the two most commonly used scoring systems are the Glucksberg grading‐ (Grades I‐IV) and the Center for International Blood and Marrow Transplant Research (CIBMTR, formerly IBMTR) grading system (Grades A‐D). Both grading systems have been prospectively validated and considered equally performing. While GVHD grading does correlate with overall survival (OS), poorly surviving patients are found among individual low‐grade GVHD patients. Despite such limitations, no superior scoring system had gained universal acceptance since. Recent GVHD severity scores involved additional concepts, such as epithelial damage in GVHD or predictive biomarkers.Aims:Machine Learning (ML) models have the potential to class and predict events in transplantation medicine. We aimed to improve GVHD grading using ML models in order to improve survival prediction within GVHD scoring.Methods:We analyzed the Glucksberg scoring system in 1354 consecutive adult patients with GVHD after HSCT from a single center between 2008 and 2018 and compared it to a ML model, based on the available GVHD organ involvement staging, transplant and OS data. We constructed a 3D discrete space (V), of which axes corresponded to the acute GVHD stage of the corresponding organ (skin, liver, intestine). Each individual patient's GVHD stages defined a point in V. A principal component analysis applied on the data and the first principal component was adopted as the new space (V ∗ ) of the SI. Then, by using the first principal component we formulated an algebraic relation to obtain a SI of each patient. The performance of the ML and Glucksberg grading systems were compared within Cox regression survival analysis.Results:The ML model revealed that GVHD grading according to the Glucksberg grading was an intact model that still insufficiently correlated with OS (Figure). We then constructed a SI with a stronger differentiation capacity with respect to OS. Cox regression survival analysis compared the proposed ML grading system to the Glucksberg grading using the SI as a risk factor and showed that the ML‐based SI offered a finer grading (125 grades, allowing to use the SI as a continuous variable). The ML model was more stable with respect to censoring period than classical grading. We observed greater variations of the Glucksberg grading, which changed over different censoring periods from a HR of 1.79 (95% Confidence interval, 1.61–1.99) to 1.56 (95%CI, 1.44–1.70), while in the ML‐grading the HR remained stable with a small CI (HR 1.55; 95%CI, 1.45–1.66). The ML‐based SI allowed to model individual patient's maximum survival expectation. Finally, the suggested ML algorithm offered a better predictive factor, because it created cohorts with more distinct hazard rates.Summary/Conclusion:A ML based scoring system using the Glucksberg organ staging categories allowed refined GVHD grading and at the same time improved OS prediction across different GVHD severity cohorts.image