The simplified age and stage (SAS) score improves OS prediction after allogeneic stem cell transplantation

Abstract Background Allogeneic hematopoietic cell transplantation (allo-HCT) is a potentially curative treatment for acute leukemia and myelodysplastic syndromes (MDS), yet long-term survival remains limited. Currently established prognostic scores differ in complexity and predictive accuracy. We aimed to systematically benchmark conventional prognostic models and develop a simplified, data-driven score to improve overall survival (OS) prediction after allo-HCT. Methods We retrospectively analyzed 561 adult allo-HCT recipients treated at our institution between 2010 and 2024. Of these, 421 patients with complete data on EBMT Score, HCT-CI Score, EASIX Score, log₂EASIX, ECOG Score, and OS formed the primary analysis cohort. We first benchmarked existing prognostic scores using Harrell's concordance index (C-index). A simplified model was then developed in an expanded cohort (n = 525; 421 patients from 2010–2022 plus 104 from 2023–07/2024) using forward variable selection in a Cox proportional hazards framework. All analyses were performed using reproducible Python code. Results Among conventional scores, the EBMT Score performed best (C-index 0.586), while HCT-CI, EASIX, log₂EASIX, and ECOG showed lower discrimination (C-index 0.445–0.550). The final model included age at transplant (in years) and disease stage, categorized according to the original EBMT Score definitions, and weighted by Cox regression coefficients. The resulting simplified Age and Stage (SAS) Score was calculated as: SAS Score=Age+9 × (Disease Stage) Internal validation using 1000-fold bootstrapping showed that the SAS Score significantly outperformed the original EBMT Score, with an absolute C-index improvement of +0.045 (95% CI: 0.0131–0.0787). Likelihood ratio testing also confirmed superior model fit (LR = 20.06; p = 7.5×10⁻⁶). In addition, SAS Score demonstrated better calibration (intercept: 0.026 vs. 0.195; slope: 0.533 vs. 0.302) and comparable overall prediction error (Brier score at 1 year: 0.280 vs. 0.257). In comparison, a logistic regression using the same variables achieved similar discrimination (C-index 0.613, AUC: 0.613), while a random forest model showed signs of overfitting (C-index 0.865, AUC: 0.497). Stratification by SAS Score tertiles (low: ≤58.1, intermediate: 58.2–71.8, high: ≥71.9) revealed a significant OS difference between high- and low-risk groups (log-rank p < 0.00001), supporting the model's discriminative capacity. Conclusion We developed a transparent, statistically robust, and clinically applicable prognostic score using only two variables. Compared to existing models, the SAS Score improves OS prediction and enables effective risk stratification. It may support pre-transplant decision-making and warrants prospective validation in multicenter settings.