Unveiling CMV Risk Signatures After Allogeneic Stem Cell Transplantation Through Machine Learning Approaches: A Multicenter Analysis From the GETH-TC.

Cytomegalovirus (CMV) infections remain a major complication following allogeneic hematopoietic stem cell transplantation (allo-HCT), with refractory infections associated with increased morbidity and mortality. Machine learning (ML) approaches may improve risk stratification and clinical management of CMV infections. This retrospective multicenter study analyzed 933 allo-HCT recipients from 21 Spanish centers (2014-2018) using data from the Spanish Hematopoietic Transplantation and Cell Therapy Group registry. Three ML models were developed to predict: (1) CMV DNAemia occurrence, (2) clinically significant CMV infection (csCMV-I), and (3) refractory CMV infection (CMV-R). Six supervised ML algorithms were evaluated using 25 repeated fivefold stratified cross-validation, including Random Forest, Support Vector Machine, Decision Trees, XGBoost, Elastic Net, and PLS-Logistic Regression. CMV DNAemia occurred in 493 recipients (53%), with 356 (72%) developing CsCMV-I and 91 (26%) progressing to CMV-R. The optimal models achieved: XGBoost for CMV DNAemia and CsCMV-I prediction (sensitivity 0.97) and SVM for CMV-R prediction (sensitivity 0.58). Key predictive features included recipient CMV serostatus, conditioning regimens, GvHD occurrence, recipient characteristics, and immunosuppressive protocols, with distinct signatures for each outcome. ML models successfully identified distinct clinical profiles associated with different CMV outcomes in allo-HCT recipients, achieving high sensitivity but moderate discriminatory capacity (AUC 0.55-0.67). Despite limitations including pre-letermovir era data, this work establishes a foundation for precision risk stratification as a complement to-not replacement for-current prophylaxis strategies. Prospective validation in contemporary cohorts is needed before clinical implementation.