Real-world temporal evaluation of static risk prediction in polycythemia vera

Abstract Current prognostic models for polycythemia vera (PV)—including ELN (Barbui et al. Leukemia, 2018), MIPSS-PV (Tefferi et al. Br J Haematol, 2020) and the Myeloproliferative Neoplasm Personalized Risk Calculator (MPN-PRC) (Grinfeld et al. NEJM, 2018)—were all developed using diagnostic (dx) data to guide initial treatment. In practice, however, they are often reapplied during follow-up to inform ongoing care. Dynamic application has not been validated and may misrepresent evolving risk. We evaluated model performance over time in a real-world PV cohort. We analyzed 475 patients (pts) with PV treated at our institution (median age at dx 56 years [yrs], median follow-up 11 yrs) (Abu-Zeinah et al. Leukemia, 2021). Of 225 pts with complete genomic data, 53% had ≥1 co-occurring mutation, and 19% had an abnormal karyotype. Outcomes were estimated using Kaplan-Meier (KM) methods, and multivariable Cox proportional-hazards models were used to define associations. The ELN model, based on age ≥60 or thrombosis history (hx), stratifies thrombosis-free survival (TFS) at dx (p=0.00037 in our cohort). We reapplied ELN every 5 yrs post-dx, tracking risk shifts, new thrombotic events, and hazard ratios (HRs). Most pts shifted into the high-risk group due to aging. By 5 yrs, ELN parameters lost prognostic significance (HRs p>0.05), and 10-year TFS converged across risk groups. These findings highlight the importance of using models within the temporal scope for which they were developed. MIPSS-PV stratified overall survival (OS) at dx (p<0.0001), but in our cohort, age dominated mortality risk (age >67 yrs HR 6.3 [1.6–5.6], p<0.001) while WBC ≥15, SRSF2 mutations, and thrombosis hx were not significant. When reapplied every 5 yrs, MIPSS-PV progressively reclassified pts into higher risk groups, yet 10-yr OS converged across strata. This identifies a disconnect between static risk frameworks and dynamic disease biology. Models must be applied for their intended outcomes. ELN was developed for TFS, not OS or myelofibrosis-free survival (MFS); MIPSS-PV for OS, not other outcomes. In our cohort, ELN modestly stratified OS and MFS at dx, but this was entirely driven by age (OS: age ≥60 HR 8.9 [5.7–14], p<0.001; MFS: HR 2.1 [1.3–3.4], p=0.002). MIPSS-PV stratified TFS and MFS weakly via age alone and performance declined over time. These findings illustrate the risk of applying models beyond their designed scope and the dominant role of age in predicting OS. To evaluate the MPN-PRC in PV, clinical/genomic data were uploaded to the online calculator (Grinfeld et al. https://www.sanger.ac.uk/tool/progmod/progmod/) to generate pt-specific predictions for event-free survival (EFS) and OS, MFS, and leukemia-free survival (LFS) at 5-, 10-, and 20-yrs post-dx. Actual outcomes were estimated using KM methods and compared to predictions using truncated c-indexes (Uno et al. Stat Med, 2011; Shi, Github, https://github.com/YushuShi/correctedC)with 1000 bootstraps. The MPN-PRC stratified OS well at all timepoints (c-index ≥0.80). MFS prediction was strong at 5 yrs (c-index 0.91) but deteriorated at 10/20 yrs (c-index ~0.5). LFS prediction was poor (c-index <0.7), likely due to low leukemic transformation in our cohort (10-yr LFS = 99%). EFS was consistently overestimated (r = 0.43). To identify the variables driving MPN-PRC predictions, we systematically varied age, sex, and blood counts across observed ranges, adding individual co-mutations. Outputs were bootstrapped 10,000 times. Age dominated OS predictions; high WBC and TP53 and SRSF2 mutations modestly increased 10/20 yr post-dx risk. For MFS, SRSF2 had the strongest effect, with high PLT adding mild risk. For LFS, TP53 and SRSF2 most strongly increased predicted AML risk. We conclude: 1. Prognostic models should be used within scope: ELN, MIPSS-PV, and MPN-PRC stratify PV risk effectively at dx but reapplying them beyond 5 yrs fails to reflect evolving risk. These tools remain valuable when used for their validated outcomes and timeframes. Overextending scope can mislead clinical decisions. 2. Age dominates OS prediction: Age is immutable and is the strongest predictor of OS. Any new prognostic variable must be benchmarked against an age-only model to establish added value. 3. Future models must be dynamic: Therapies and life events reshape disease trajectories. PV risk assessment must move from static snapshots to time-sensitive, adaptive frameworks that reflect the continuous nature of PV care and risks.