Abstract Introduction Prognostic models like the IPSS-R and IPSS-M focus on the biology of the disease but ignore patient-related factors such as frailty and comorbidities, which are essential for clinical outcomes. We aim to develop integrated models that integrate both to improve prognosis assessment. Methods A total of 687 consecutive patients (pts) with newly diagnosed de novo MDS who visited our center from 08/2016 to 06/2023 were included. Clinical and biological characteristics were recorded at diagnosis. Pts were randomly divided into a training cohort (80%) and a test cohort (20%). Variables were selected both by variable importance (VIMP) and minimal depth. A machine learning technique, random survival forests (RSF) was used to develop the prognostic model. Five-fold cross-validation (CV) was used both for hyperparameter tuning and for internal validation of the model in the training cohort. The final model was then applied to the test cohort for external validation. The performance of the model was evaluated using Harrell's Concordance Index (C-index), Brier scores and area under the curve (AUC) of time-dependent receiver operator characteristic (ROC) curve compared with established prognostic models. Results The IPSS-R integrated model was constructed using 7 variables selected by the algorithm: IPSS-R, pulmonary, cardiac and cerebrovascular disease, hypertension, self-care ability, and solid tumor. It achieved a C-index of 0.700/0.708 and Brier scores of 0.175/0.140 in the training/test sets, respectively. According to the IPSS-R integrated model, three risk-groups were defined: low (38.0%), intermediate (35.5%), and high-risk (26.5%), with a median OS of unavailable, 37.4 (95% CI, 29.5-45.4) and 12.8 months (95% CI, 10.7-15.0), respectively(p0.001). The 1-/3-/5-year OS rates were 90.2%/78.7%/65.3%, 82.6%/63.2%/27.9%, and 55.3%/28.2%/10.2%. AUCs at 1/3/5 years were 0.713/0.768/0.713, outperforming IPSS-R (0.688/0.714/0.660) and comparable to IPSS-M (0.717/0.754/0.700). Subgroup analyses across age, treatment types, and BM blast percentages confirmed the model's robustness (p0.001). Among patients classified as very low- or low-risk by IPSS-R, those up-staged by the integrated model (n=49) had shorter OS (36.7 months vs. not reached; p0.001). Among high- or very high-risk patients, those down-staged (n=148) had longer OS (37.7 vs. 12.8 months; p0.001). We then constructed an IPSS-M integrated model combining molecular and patient-related factors, achieving AUCs of 0.759/0.806/0.755 at 1/3/5 years. Four risk-groups were defined: low (25.1%), intermediate-1 (25.3%), intermediate-2(24.3%) and high-risk (25.1%), with a median OS of unavailable, 59.0 (95% CI, 40.2-77.8), 27.6 (95% CI, 21.9-33.3) and 11.8 months (95% CI, 9.5-14.1), respectively(p0.001). Subgroup analyses confirmed consistent performance. Among patients initially classified as very low-risk or low-risk by IPSS-M, those up-staged (n=30, median OS 36.7 months) by the IPSS-M integrated model demonstrated significantly shorter OS compared with unchanged patients (n=105, median OS unavailable, P0.001). Notably, 60.0% of these up-staged patients were elderly individuals aged 60 years. Among patients classified as high- or very high-risk, those down-staged (n=95, median OS 54.7 months) by the IPSS-M integrated model demonstrated significantly better OS compared with unchanged patients (n=163, median OS 11.8 months, P0.001). Moreover, in IPSS-M high-/very high-risk groups treated with hypomethylating agent treatment, the overall response rate was significantly higher in down-staged patients (17, 73.9%) compared to unchanged patients (47.6%, P=0.041), and the incidence of febrile neutropenia with infection was higher in unchanged patients (37, 72.5%) than down-staged patients (42.9%, P=0.017). Conclusion Integrating patient-related factors into IPSS-R improved risk stratification and accuracy to a level comparable with IPSS-M, providing a practical alternative where molecular testing is limited. The integration of molecular data with patient-related factors improves the accuracy of prognostic models and has important clinical implications, particularly in the elderly population.
Wang et al. (Mon,) studied this question.