Amer Zeidan
Yale University,USA
Title: The utility of risk models in predicting outcomes of patients (pts) with higher-risk myelodysplastic syndromes (HR-MDS) treated with hypomethylating agents (HMA)
Biography
Biography: Amer Zeidan
Abstract
Background: Although HMA (azacitidine [aza] or decitabine [dac]) are standard of care therapies for pts with HR-MDS, responses may not be seen for 4-6 months and occur in <50% of treated pts. Moreover, curative hematopoietic cell transplantation (HCT), which is recommended as up-front therapy for HR-MDS, is used USA in <5% of pts. The ability to identify pts with a low likelihood of benefiting from HMA who instead should receive immediate HCT or experimental treatment approaches is a clinical and research priority. Established MDS prognostic systems include the International Prognostic Scoring System (IPSS) and the revised IPSS (IPSS-R) which were derived from untreated pt cohorts and the MD Anderson Prognostic Scoring System (MDAPSS) which was derived from a cohort composed of treated and untreated pts. A French Prognostic Scoring System (FPSS) was developed specifically to predict survival benefit among aza-treated HR-MDS pts. We sought to compare the relative prognostic discriminatory power of these models in a large cohort of HMA-treated pts with HR-MDS. Methods: The combined MDS database obtained from six institutions in the MDS Clinical Research Consortium (H. Lee Moffitt Cancer Center and Research Institute, Cleveland Clinic, MD Anderson Cancer Center, Dana Farber Cancer Institute, Weill Medical College of Cornell University, and the Sidney Kimmel Comprehensive Cancer Center at Johns Hopkins) was used to identify patients with HR-MDS (IPSS intermediate-2 [INT-2] and high) who received HMA therapy (aza or dac). The prognostic scores were calculated as previously described. Responses were defined per International Working Group 2006 criteria. According to best response, pts were categorized into responders (complete response [CR], partial response [PR], hematologic improvement [HI], and marrow CR) and non-responders (stable disease [SD] and progressive disease [PD]). Multiple imputation using the chained equation approach was used to impute all missing data. 100 imputations were generated from a model with 82 variables using random forest imputation for missing continuous data and logistic/polytomous imputation for missing categorical data. All statistical analyses were performed on the 100 completed datasets, then combined using the standard multiple imputation combining rules. Fraction of missing information (FMI), which gives the fraction of information that was lost for a quantity of interest due to missing data under a given imputation model, was estimated for every analysis from the 100 imputations. Logistic regression models were fitted and tested for association of response with prognostic risk categories. Overall survival (OS) was calculated from the time of diagnosis to time of death or last follow-up. Kaplan-Meier (KM) curves were generated for OS and the stratified (by institution) log-rank test was used to compare median OS. Stratified (by institution) Cox proportional hazards models were fit to assess association of prognostic systems with OS. Corrected Akaike information criteria (AICc) were used to assess the relative goodness of fit of these fitted models. Results: We identified 626 pts with HR-MDS (69.9% with INT-2 and 30.1% with high IPSS) who received HMA as upfront therapy (67.8% aza, 32.2% decitabine). Median duration of follow-up from diagnosis was 15.5 months (M) (95% confidence interval [CI], 14.5-16.7 M): 66.4% of pts were male, 86.6% white, 85.0% were > 60 years, and 8.9% had therapy-related MDS. Median number of HMA cycles was 4.4 (1st-3rd Quartiles, 3.0-7.6), with 70.8% of pts receiving ≥4 cycles of therapy. Median time from diagnosis to start of HMA was 0.94 M (1-3 Q, 0.40-2.1 M). In total, 43.0% responded (CR 20.7%, PR 9.5%, HI-E 6.4%, HI-N 3.6%, HI-P 2.7%) while 57.0% were non-responders (SD 37.9% and PD 19.1%). Consistent with literature, none of the prognostic systems predicted best response to HMA: IPSS (P=0.55), IPSS-R (P=0.14), FPSS (P=0.21), MDAPSS (P=0.39), WPSS (P=0.88). Median OS for the entire cohort was 16.9 M (95%CI, 15.6-18.2 M, FMI=3.3%). Figure 1 shows the KM curves by the 5 prognostic models for a representative completed dataset. All prognostic models showed association with OS: IPSS (P=0.036), IPSS-R (P=1.6e-10), WPSS (P=0.00020), FPSS (P=4.0e-9), MDAPSS (5.7e-13). Scores generated using the AICc to assess the relative goodness of fit (lower is better) were 4146 (MDAPSS), 4149 (FPSS), 4157 (IPSS-R), 4189 (WPSS) and 4207 (IPSS). Conclusions: This is the largest reported direct comparison of the commonly used prognostic models for MDS among HMA-treated MDS patients. None of the prognostic models predicted best response to HMA therapy. Nonetheless, the IPSS-R, MDAPSS, and the FPSS all functioned well to separate HMA-treated pts with HR-MDS into prognostic groups with different survivals. The MDAPSS and FPSS appear superior to the IPSS-R, WPSS and the IPSS for survival prediction among HMA-treated pts. HR-MDS patients with poor projected survival with HMA therapy should be considered for up-front HCT or for experimental approaches.
