Artificial Intelligence (AI) Based, Machine Learning (ML) Predicting the Individual Absolute Risk of Acute Graft Versus Host Disease (aGvHD) in a Retrospective International Cohort

Background: HSCT still holds potentially lethal complications such as aGvHD, relapse of malignant disease, graft rejection, and infections. 30 to 50% of all patients undergoing HSCT will develop aGvHD, causing high morbidity and mortality (grade 4 GvHD up to 90%) as well as generating significantly higher treatment costs than a HSCT without aGvHD. MatchGraft.AI is an AI based ML algorithm that predicts individual, absolute risk of aGvHD for every individual HSCT donor-recipient pair, based on data available prior to start of conditioning aiming for optimization of donor selection.

Methods: MatchGraft.AI has been built in accordance with best clinical practice (Higgins et al 2021). A random forest is used to predict the probability of aGvHD free survival, GvHD free death and aGvHD. Missing data are imputed, except for the gradient boosted tree, by using the most frequent category for categorical variables and mean value for continuous variables. The performance characteristics are calculated in an aGvHD vs Rest (death or aGvHD free survival) fashion. The data is split into a stratified training and validation set with a ratio of 80% / 20%.

Results: The model was trained at Charité University against a retrospective, multi center data set of 1035 German and north American patient/donor sets. We collected 1719 data sets. 684 cases were excluded due to mayor outcome parameters not being available. The following analyzation focuses on the 1035 included stem cell transplantation patient/donor sets (828 training set/ 207 validation set), 356 pediatric patients, 679 adults. M:F = 1,5; age 0 – 76 (median: 39.5 +/- 24.3 years). Indications for transplantation were malignant diseases (hematological neoplasms, lymphomas, solid tumors) in 731 patients and benign diseases (hemoglobinopathies, immune deficiencies, bone marrow failure) in 205 cases, in 99 cases diagnosis was not documented. Transplantations were performed between January 2013 and December 2019. Despite some missing data in our retrospective cohort, model performance was estimated at 0.70 AUC (fig.1). Integrated calibration index was 0,0277.

Discussion: MatchGraft.AI predicts the risk of acute GvHD. By providing precise prediction of relevant complications, we expect the pool of possible donors to become larger as well as donor selection to improve. Furthermore, physicians and patients will be able to better prepare and adapt conditioning and treatment plans with MatchGraft.AI predictions.

Conclusions: The prediction of individual aGvHD risk, using this AI ML model, is feasible and well calibrated, especially in the medium risk area. Ongoing continuous training, testing and adapting on additional international, multicenter retrospective and prospective data will further improve performance in the low risk area. This study is the basis for further prospective research.