Post-transplant lymphoproliferative disorder (PTLD) is a potentially life-threatening complication that occurs in approximately 1%–16% of solid organ transplant (SOT) recipients [1–4]. The intensity of immunosuppression and Epstein–Barr virus (EBV) replication are key drivers in the development of PTLD [1, 5]. The risk of PTLD is particularly high in EBV-seronegative recipients who acquire primary EBV infection post-transplant, especially when the donor is EBV-seropositive, highlighting the pivotal role of EBV serostatus in PTLD development [6]. The association between cytomegalovirus (CMV) infection and PTLD remains inconclusive, with inconsistent findings across studies. Some studies have identified CMV disease and CMV-EBV coinfection as risk factors for PTLD, while others found no consistent association after adjusting for EBV serostatus and immunosuppression intensity [7–12]. The association between CMV infection and post-transplant outcomes, including PTLD, may be confounded or modified by antiviral prophylaxis [13, 14]. Although some data showed antiviral agents such as valganciclovir may delay EBV viremia, the role of antiviral prophylaxis in preventing PTLD remains debated [15, 16]. A systematic review found insufficient evidence to support this conclusion [17–19]. However, a recent meta-analysis suggests antiviral prophylaxis may reduce PTLD risk and EBV viremia, especially in patients receiving intensive immunosuppression [20].

Inconsistencies in prior observational studies evaluating PTLD risk likely reflect methodological limitations, including small sample sizes, short follow-up durations, single-center designs, and lack of adjustment for CMV serostatus. In this study, we applied machine learning (ML) to estimate PTLD risk in SOT recipients, incorporating EBV-CMV serostatus as a compound variable to explore whether patients with [EBV (D+/R−), CMV (D−/R−)] serostatus are at elevated risk of PTLD.

During the study period, a total of 47,333 patients underwent abdominal organ transplants (median [IQR] age: 49 [37–58] years, male: 28,718 [60.6%]). Cohort characteristics are provided in Table 1.

EBV-naïve recipients are exposed to donor-derived EBV infection and are vulnerable to PTLD due to unregulated viral replication under immunosuppressive conditions [1]. Previous studies consistently showed patients with EBV D+/R− serostatus are at the greatest risk of PTLD development [25–27]. However, the relationship between CMV serostatus and PTLD has not been thoroughly investigated. Using two ML models, we found a consistent association between [EBV [(D+/R−), CMV (D−/R−)] serostatus and increased PTLD risk. Among EBV mismatch recipients, those with CMV D+/R− had a significantly lower risk of PTLD compared to those with CMV D−/R−. Although single-center studies have suggested a potential link between PTLD and CMV infection, the association between PTLD and CMV serostatus has not been systematically investigated [10, 28, 29]. An early study by Walker et al. in 1995 examined a cohort of 381 non-renal SOT recipients at the Mayo Clinic and proposed a possible role for CMV serostatus in PTLD pathogenesis. However, the interpretation of these results is constrained by confounding from concurrent use of potent immunosuppressive therapy, notably OKT3, which substantially elevates PTLD risk [9]. Opelz et al. conducted a large SRTR-based cohort study (1999–2007) of 23,340 SOT recipients and found no significant association between recipient CMV serostatus and non-Hodgkin lymphoma; however, donor CMV serostatus was not included in their analyses [30]. In a separate cohort study, Geris et al. found a relationship between CMV D-/R- serostatus and the development of PTLD. However, the study outcome was limited to diffuse large B-cell lymphoma (DLBCL) and did not investigate the simultaneous roles of EBV (D/R) and CMV (D/R) as a compound variable [7]. Ali et al., in a cohort of 10,947 pediatric renal transplant recipients, found that recipient CMV seropositivity was associated with a protective effect against PTLD (HR = 0.82; 95% CI: 0.73–0.94) [11]. To our knowledge, no prior study has evaluated the association between PTLD and combined EBV/CMV donor–recipient serostatus as a single compound variable.

We evaluated PTLD risk using combined [EBV (D/R), CMV (D/R)] serostatus and found that EBV D+/R- recipients with CMV D−/R− serostatus were at increased risk of PTLD. This observation does not inherently indicate a protective effect of valganciclovir, which is typically not administered to CMV D-/R- recipients. Although they may receive acyclovir or valacyclovir, these agents have negligible activities against CMV and no proven benefit in preventing EBV-related PTLD. Aldabbagh et al. conducted a metanalysis in 2017 including 9 studies and 2,366 SOT recipients and concluded that existing data are insufficient to support the routine use of antivirals to prevent PTLD in EBV-naive SOT recipients [19]. However, a recent meta-analysis by Moghadamnia et al., including 22 studies and 13,498 SOT recipients, found that PTLD incidence was significantly lower among those who received antiviral prophylaxis (RR 0.77; 95% CI, 0.63–0.94) [20]. In heart and kidney transplant recipients, Albatati et al. found that valganciclovir delayed viremia onset compared to no antiviral use (143 vs. 90 days; p = 0.008), with each additional day of prophylaxis increasing viremia-free survival by 1.4% (p < 0.001) [15]. In a case-control study of 100 PTLD cases and 375 controls, ganciclovir use was associated with a 38% reduction in early PTLD risk for each 30-day period during the first posttransplant year [17]. Lytic replication during primary EBV infection promotes dissemination via infected B cells and remains susceptible to antiviral suppression [31]. However, these findings do not necessarily support the use of nucleoside analogs in EBV D+/R- recipients, who are at the highest risk of primary infection. Antiviral prophylaxis is not currently recommended by current guidelines, as evidence remains limited [1]. In the absence of interventional studies evaluating preventive strategies, further research is warranted to identify approaches that reduce PTLD risk.

In our study, 60% of PTLD cases occurred in male transplant recipients, consistent with prior research indicating the disproportionate impact of PTLD, particularly late-onset disease, on males [30]. This disparity may stem from differential immune responses and increased susceptibility to EBV-associated complications [14, 32]. We also observed that patients with PTLD were older than those without, likely reflecting increased risk due to immunosenescence [33]. Most PTLD cases in our cohort were late-onset (median time from transplant to diagnosis: 6.1 years), consistent with prior studies linking late-onset PTLD to older age and early-onset PTLD to younger recipients [30].

We found an inconsistent association between PTLD and rabbit antithymocyte globulin (rATG) use, despite previous studies showing a possible link [34, 35]. This inconsistency may be partly explained by selective survival bias: recipients at a higher risk for PTLD are less likely to receive rATG, as clinicians frequently avoid its use in these patients due to its known association with PTLD [36, 37]. Additionally, 88% of patients with PTLD in this cohort had a late-onset disease, typically many months or even years after rATG exposure.

The application of newer ML models underscores the evolving methodology in estimating PTLD risk. RFC particularly aggregates predictions from multiple decision trees to mitigate overfitting and enhance generalizability [38, 39]. These strengths make RFC particularly suitable for analyzing highly imbalanced datasets and predicting rare events [40]. Previous studies have demonstrated that RFC often outperforms LR in terms of predictive accuracy and interpretability in various medical domains, including organ transplantation [41, 42]. In our study, RFC and LR models produced comparable AUC values of approximately 0.65 in the training set and 0.61 in the testing set. Model performance was likely impacted by extreme class imbalance with PTLD occurring in only 1.5% of patients. To address this, we employed random undersampling repeating the process 20 times to balance the dataset and improve model sensitivity to minority-class predictions [21]. Although random undersampling risks excluding valuable majority-class data, this iterative approach helped achieve a balance between data retention and model focus [43]. We could not include granular clinical, dynamic, or omics data—such as medication dosages, viral load, or immune markers—which likely limited model performance. These individual-level variables are unavailable in population-based sources like the SRTR. The modest overall performance may also reflect the limited granularity of the feature set, which may not adequately capture individual-level variation in PTLD risk. Incorporating personalized variables, such as pharmacogenetic or molecular data, could improve future model performance.

Our study had some limitations. First, the retrospective cohort design is subject to inherent constraints, including the potential for selection bias. Due to considerable missing data on EBV and CMV serostatus among thoracic transplant recipients, we excluded this subgroup from the cohort. Another limitation is the small number of PTLD cases, which reduced model performance; as such, our analysis is intended to generate hypotheses and highlight potential risk patterns rather than serve as a clinical prediction tool. The predictive performance of both logistic regression and random forest models was modest (test AUC ∼0.61), which likely reflects several methodological challenges: the low incidence of PTLD, the resulting class imbalance despite undersampling, and the absence of granular clinical variables such as viral load, dynamic immune monitoring, heterogeneity in immunosuppressive regimens, donor-recipient characteristics, and center-specific protocols in the SRTR database. These limitations may have reduced the ability of our models to achieve higher discrimination.

This study demonstrates the utility of ML models in predicting the rare post-transplant outcomes such as PTLD in the SOT population. Our study may challenge traditional reliance on linear models. By bridging established research with cutting-edge methodologies, this study paves the way for improved clinical decision-making and personalized patient care. Our findings have potential implications in the PTLD risk assessment and suggest using compound [EBV (D/R), CMV (D/R)] serostatus rather than simple EBV (D/R) for PTLD prediction.