EXPLORING A NEW PROGNOSTIC INDEX IN 341 POST-TRANSPLANT LYMPHOPROLIFERATIVE DISORDERS IN THE “RITUXIMAB ERA”: A RETROSPECTIVE STUDY FROM SPANISH LYMPHOMA GROUP GELTAMO

D. Gil-Alós and A. Jiménez-Ubieto equally contributing author.

Introduction: Post-transplant lymphoproliferative disorders (PTLD) are life threatening malignancies that can develop as a complication following solid organ transplantation. Due to the risk of treatment-related complications, tailoring treatment to patients according to their risk profile could potentially improve outcomes even further. The aim of this study was to design a simple and practical prognostic model and to validate scores usually used in large b cell lymphoma in a large cohort of monomorphic PLTD patients.

Methods: A total of 341 PTLD patients from 16 Spanish hospitals (2000–2024) were retrospectively included. 27 clinical and biological variables were collected, excluding 2 with > 35% missing data. The remaining missing values (∼7%) were imputed using KNN. Feature selection was performed using LASSO, followed by further refinement with the Elbow method of Shapley values, retaining the top 80% of cumulative feature importance. A random survival forest model was trained to develop the Artificial Intelligence-PTLD Prognostic Index (AI-PTLD-PI), using an 80/20 train-test split. The training underwent 5-fold cross-validation for hyperparameter tuning, followed by final model retraining and evaluation on the independent test set. The AI-PTLD-PI was compared with established scores (IPI, aaIPI, R-IPI, NCCN-IPI, GELTAMO-IPI, KPI, DLBCL-IPI, PTLD-IPI) using time-dependent AUC, C-index, and Brier score at 5 years.

Results: Among the patients analyzed, 85% had B-cell monomorphic PTLD, mainly following kidney (42%) or liver transplantation (30%). The median time from transplant to lymphoma onset was 91 months (range: 1–393), and 41% of cases were EBV-related. The median follow-up was 5.7 years.

To assess prognostic accuracy, univariate Cox regression was applied to each index. The modified NCCN-IPI and PTLD-IPI showed the highest performance among standard scores (c-index: 0.672 and 0.639, respectively). Our AI-PTLD-PI outperformed all models, achieving a c-index of 0.736. In time-dependent analyses, AI-PTLD-PI consistently demonstrated superior prognostic value with a time-dependent AUC of 0.783. Detailed performance comparison is shown in Figure 1A,B.

For further validation, AI-PTLD-PI stratified patients into survival terciles (low, intermediate, high risk) (Figure 1C). AI-PTLD-PI showed a log-rank p-values of 0.09 (low vs. intermediate), 0.000002 (low vs. high), and 0.001 (intermediate vs. high), reinforcing its prognostic utility. Key prognostic factors in AI-PTLD-PI were found with a Shapley value analysis, highlighting older age at transplant and diagnosis, lower albumin, elevated LDH, and high lymphocyte count as major risk factors (Figure 1D).

Conclusion: In this large PTLD series, the AI-PTLD-PI identify patients at higher risk of death with superior accuracy than other scores. The major risk factors are based on age plus three variables routinely evaluated at diagnosis through a basic blood test: LDH, albumin, lymphocytes.

Keywords: aggressive B-cell non-Hodgkin lymphoma; diagnostic and prognostic biomarkers

No potential sources of conflict of interest.