AUTHORS
Yue Lyu, PhD, Kelly Sharon Chien, MD2, Koji Sasaki, MD, PhD3, Ian Bouligny, M.D.4, Jing Ning5, Xuelin Huang, PhD6, Alex Bataller, MD, PhD7, Alexandre Bazinet, MD, MSc8, Wei Ying Jen, BM BCh, FRCPath9, Steven Hsesheng Lin, MD10, Mahesh Swaminathan, MD11, Danielle Hammond, MD12, Samuel Urrutia, MD, MS13, Juan Jose Rodriguez-Sevilla14, Guillermo Montalban-Bravo, MD15, Guillermo Garcia-Manero, MD16, Ziyi N/A Li17
AFFILIATIONS
1 The University of Texas MD Anderson Cancer Center, Houston, TX, United States
2 The University of Texas MD Anderson Cancer Center, Houston, TX, United States
3 The University of Texas MD Anderson Cancer Center, Houston, TX, United States
4 The University of Texas MD Anderson Cancer Center, Houston, Texas, United States
5 The University of Texas MD Anderson Cancer Center, Houston, TX, United States
6 The University of Texas MD Anderson Cancer Center, Houston, TX, United States
7 The University of Texas MD Anderson Cancer Center, Houston, Texas, United States
8 The University of Texas MD Anderson Cancer Center, Houston, Texas, United States
9 The University of Texas MD Anderson Cancer Center, Houston, Texas, United States
10 The University of Texas MD Anderson Cancer Center, Houston, Texas, United States
11 The University of Texas MD Anderson Cancer Center, Houston, Texas, United States
12 The University of Texas MD Anderson Cancer Center, Houston, Texas, United States
13 Washington University in St Louis, St Louis, Missouri, United States
14 Hospital Clínic de Barcelona, Barcelona, Spain
15 The University of Texas MD Anderson Cancer Center, Houston, Texas, United States
16 The University of Texas MD Anderson Cancer Center, Houston, Texas, United States
17 The University of Texas MD Anderson Cancer Center, Houston, TX, United States
ABSTRACT
Background: Prognostic models for myelodysplastic syndromes (MDS), including the Revised International Prognostic Scoring System (IPSS-R) and Molecular International Prognostic Scoring System (IPSS-M), classify risk using data collected at a single time point, although prognosis evolves during follow-up. We developed the Dynamic Prognostic Scoring System (DPSS), a landmark-based SuperLearner framework that integrates up-to-date clinical, cytogenetic, and molecular data to generate individualized, time-updated survival predictions for patients with MDS.
Methods: We identified 1,085 patients in a training cohort and 317 patients in a temporally independent validation cohort from a tertiary cancer center. Among these, 878 and 259 patients with complete data were included in model development and validation, respectively. DPSS used prespecified landmark times at 0, 6, 12, 18, and 24 months after diagnosis. At each landmark, only patients who remained alive and under observation were included, and predictors were defined using the most recent information available before or at that landmark. Predictors included age, sex, hemoglobin concentration, platelet count, absolute neutrophil count, bone marrow blast percentage, IPSS-R cytogenetic risk category, and mutational status of 10 recurrent genes. Five-fold cross-validation was used to train a SuperLearner ensemble incorporating Cox proportional hazards regression, random survival forests, accelerated oblique random survival forests, generalized boosted regression models, and eXtreme Gradient Boosting. Predictions were combined using weights chosen to minimize inverse-probability-of-censoring-weighted Brier loss. Discrimination was assessed using Uno's time-dependent concordance index (C-index). DPSS was compared with IPSS-R and IPSS-M when held fixed from diagnosis and when recalculated using updated data at each landmark.
Results: DPSS achieved time-dependent C-indices of 0.77-0.84 in the training cohort, consistently exceeding IPSS-M (0.64-0.78) and IPSS-R (0.64-0.75) when these scores were held fixed from diagnosis. In the validation cohort, DPSS C-indices were 0.77, 0.75, 0.78, 0.77, and 0.83 at 0, 6, 12, 18, and 24 months, respectively, compared with 0.66-0.72 for baseline-fixed IPSS-M and 0.67-0.73 for baseline-fixed IPSS-R. DPSS maintained higher discrimination when IPSS-R and IPSS-M were recalculated at each landmark using updated clinical and molecular data, with C-index differences of 0.03-0.13; the largest gains were observed at 18 and 24 months, when outcome heterogeneity was greatest and categorical models provided limited differentiation. DPSS also provided clearer Kaplan-Meier risk-group separation across all landmarks. At 24 months in the validation cohort, log-rank P values were <0.0001 for DPSS, 0.374 for IPSS-R, and 0.081 for IPSS-M. Dynamic reclassification showed that from diagnosis to 6 months, 36% of patients moved to a higher-risk group, 19% to a lower-risk group, and 44% remained in the same category; approximately 55%-60% changed DPSS risk groups at least once during follow-up. In the validation cohort, DPSS revealed additional prognostic heterogeneity within updated IPSS-M strata. Among patients classified as Moderate High by updated IPSS-M, the proportion reassigned to the Very High DPSS group increased from 2.7% at diagnosis to 7.1% at 12 months and 26.3% at 24 months. In a sensitivity analysis censoring overall survival at stem cell transplantation, DPSS maintained significant separation across all five risk strata at diagnosis, 12 months, and 24 months; among patients who did not undergo transplantation, C-indices remained 0.748-0.801. An interactive web application was developed to accept routinely available clinical, cytogenetic, and molecular inputs and generate individualized survival curves.
Conclusions: DPSS transforms routine longitudinal MDS follow-up data into individualized, time-updated survival estimates that reflect disease evolution. By integrating evolving clinical, cytogenetic, and molecular information, DPSS improves prognostic discrimination and risk-group separation beyond IPSS-R and IPSS-M, including when these scores are dynamically recalculated. DPSS may support longitudinal prognostic counseling, treatment reassessment, transplant evaluation, and clinical trial consideration.
DISCLOSURES
Yue Lyu, PhD: No relevant conflicts of interest to declare.