Machine learning algorithm as a prognostic tool for Epstein-Barr virus reactivation after haploidentical hematopoietic stem cell transplantation

Abstract

Epstein-Barr virus (EBV) reactivation is one of the most important infections after hematopoietic stem cell transplantation (HSCT) using haplo-identical related donors (HID). We aimed to establish a comprehensive model with machine learning, which could predict EBV reactivation after HID HSCT with anti-thymocyte globulin (ATG) for graft-versus-host disease (GVHD) prophylaxis. We enrolled 470 consecutive acute leukemia patients, 60% of them (n = 282) randomly selected as a training cohort, the remaining 40% (n = 188) as a validation cohort. The equation was as follows: Probability (EBV reactivation) = 11 + exp(−Y), where Y = 0.0250 × (age) – 0.3614 × (gender) + 0.0668 × (underlying disease) – 0.6297 × (disease status before HSCT) – 0.0726 × (disease risk index) – 0.0118 × (hematopoietic cell transplantation-specific comorbidity index [HCT-CI] score) + 1.2037 × (human leukocyte antigen disparity) + 0.5347 × (EBV serostatus) + 0.1605 × (conditioning regimen) – 0.2270 × (donor/recipient gender matched) + 0.2304 × (donor/recipient relation) – 0.0170 × (mononuclear cell counts in graft) + 0.0395 × (CD34+ cell count in graft) – 2.4510. The threshold of probability was 0.4623, which separated patients into low- and high-risk groups. The 1-year cumulative incidence of EBV reactivation in the low- and high-risk groups was 11.0% versus 24.5% (P < .001), 10.7% versus 19.3% (P = .046), and 11.4% versus 31.6% (P = .001), respectively, in total, training and validation cohorts. The model could also predict relapse and survival after HID HSCT. We established a comprehensive model that could predict EBV reactivation in HID HSCT recipients using ATG for GVHD prophylaxis.