Limits...
An international data set for CMML validates prognostic scoring systems and demonstrates a need for novel prognostication strategies.

Padron E, Garcia-Manero G, Patnaik MM, Itzykson R, Lasho T, Nazha A, Rampal RK, Sanchez ME, Jabbour E, Al Ali NH, Thompson Z, Colla S, Fenaux P, Kantarjian HM, Killick S, Sekeres MA, List AF, Onida F, Komrokji RS, Tefferi A, Solary E - Blood Cancer J (2015)

Bottom Line: Of note, we found that the majority of CMML patients were classified as World Health Organization CMML-1 and that a 7.5% bone marrow blast cut-point may discriminate prognosis with higher resolution in comparison with the existing 10%.Using random forest survival analysis for variable discovery, we demonstrated that the prognostic power of clinical variables alone is limited.Last, we confirmed the independent prognostic relevance of ASXL1 gene mutations and identified the novel adverse prognostic impact imparted by CBL mutations.

View Article: PubMed Central - PubMed

Affiliation: Department of Malignant Hematology, H Lee Moffitt Cancer Center and Research Institute, Tampa, FL, USA.

ABSTRACT
Since its reclassification as a distinct disease entity, clinical research efforts have attempted to establish baseline characteristics and prognostic scoring systems for chronic myelomonocytic leukemia (CMML). Although existing data for baseline characteristics and CMML prognostication have been robustly developed and externally validated, these results have been limited by the small size of single-institution cohorts. We developed an international CMML data set that included 1832 cases across eight centers to establish the frequency of key clinical characteristics. Of note, we found that the majority of CMML patients were classified as World Health Organization CMML-1 and that a 7.5% bone marrow blast cut-point may discriminate prognosis with higher resolution in comparison with the existing 10%. We additionally interrogated existing CMML prognostic models and found that they are all valid and have comparable performance but are vulnerable to upstaging. Using random forest survival analysis for variable discovery, we demonstrated that the prognostic power of clinical variables alone is limited. Last, we confirmed the independent prognostic relevance of ASXL1 gene mutations and identified the novel adverse prognostic impact imparted by CBL mutations. Our data suggest that combinations of clinical and molecular information may be required to improve the accuracy of current CMML prognostication.

No MeSH data available.


Related in: MedlinePlus

Random forest survival analysis generates a novel CMML model that is comparable to existing models. The results of the random forest survival analysis for OS (a) and LFS (b) are shown. The ROC curves for all clinical models, including the new model generated using the variables discovered with random forest analysis is shown for OS (c) and LFS (d).
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4526779&req=5

fig4: Random forest survival analysis generates a novel CMML model that is comparable to existing models. The results of the random forest survival analysis for OS (a) and LFS (b) are shown. The ROC curves for all clinical models, including the new model generated using the variables discovered with random forest analysis is shown for OS (c) and LFS (d).

Mentions: All existing CMML clinical prognostic models tested were comparable and derived using a Cox proportional hazard regression and multivariate analyses approach. To determine if a novel strategy of prognostic variable discovery could yield an improved model, we performed a random forest survival analysis. This approach iteratively bifurcates the data set based on each variable and, after over 5000 permutations, determines variables of highest importance based on their ability to successfully bifurcate CMML cases based on our desired end point of OS and LFS.24 With this approach, 23 categorical variables were considered and ranked by importance, as shown in Figure 4. The top four variables for both OS and LFS were hemoglobin level <11 g/dl, the presence of circulating blasts, a platelet count of <100 × 103/dl and an adverse karyotype as defined by the CPSS. These variables were each assigned one point, and a new prognostic scoring system was devised that stratified our CMML cases into low risk (0 points), intermediate risk (1–2 points) and high risk (3-4 points). KM survival analysis and log-rank test within our database demonstrated a significant OS difference among these risk groups at not reached (95% confidence interval: 53.6–79.2), 35.1 months (95% confidence interval: 32.6–38.4) and 13.8 months (95% confidence interval: 11.7–15.4), respectively (P<0.0001). These results, and the statistically significant differences in LFS among groups (P<0.0001), are shown in Supplementary Figure S2. Next, we tested the relative prognostic power of this novel model against other existing CMML models and found that it had the highest AUC at 0.714 for OS and second highest AUC for LFS at 0.709 (similar results for C-index). However, the difference in AUC between our novel model and existing models was not statistically significant, despite being compared within the data set for which the new model was developed (Figure 4).


An international data set for CMML validates prognostic scoring systems and demonstrates a need for novel prognostication strategies.

Padron E, Garcia-Manero G, Patnaik MM, Itzykson R, Lasho T, Nazha A, Rampal RK, Sanchez ME, Jabbour E, Al Ali NH, Thompson Z, Colla S, Fenaux P, Kantarjian HM, Killick S, Sekeres MA, List AF, Onida F, Komrokji RS, Tefferi A, Solary E - Blood Cancer J (2015)

Random forest survival analysis generates a novel CMML model that is comparable to existing models. The results of the random forest survival analysis for OS (a) and LFS (b) are shown. The ROC curves for all clinical models, including the new model generated using the variables discovered with random forest analysis is shown for OS (c) and LFS (d).
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4526779&req=5

fig4: Random forest survival analysis generates a novel CMML model that is comparable to existing models. The results of the random forest survival analysis for OS (a) and LFS (b) are shown. The ROC curves for all clinical models, including the new model generated using the variables discovered with random forest analysis is shown for OS (c) and LFS (d).
Mentions: All existing CMML clinical prognostic models tested were comparable and derived using a Cox proportional hazard regression and multivariate analyses approach. To determine if a novel strategy of prognostic variable discovery could yield an improved model, we performed a random forest survival analysis. This approach iteratively bifurcates the data set based on each variable and, after over 5000 permutations, determines variables of highest importance based on their ability to successfully bifurcate CMML cases based on our desired end point of OS and LFS.24 With this approach, 23 categorical variables were considered and ranked by importance, as shown in Figure 4. The top four variables for both OS and LFS were hemoglobin level <11 g/dl, the presence of circulating blasts, a platelet count of <100 × 103/dl and an adverse karyotype as defined by the CPSS. These variables were each assigned one point, and a new prognostic scoring system was devised that stratified our CMML cases into low risk (0 points), intermediate risk (1–2 points) and high risk (3-4 points). KM survival analysis and log-rank test within our database demonstrated a significant OS difference among these risk groups at not reached (95% confidence interval: 53.6–79.2), 35.1 months (95% confidence interval: 32.6–38.4) and 13.8 months (95% confidence interval: 11.7–15.4), respectively (P<0.0001). These results, and the statistically significant differences in LFS among groups (P<0.0001), are shown in Supplementary Figure S2. Next, we tested the relative prognostic power of this novel model against other existing CMML models and found that it had the highest AUC at 0.714 for OS and second highest AUC for LFS at 0.709 (similar results for C-index). However, the difference in AUC between our novel model and existing models was not statistically significant, despite being compared within the data set for which the new model was developed (Figure 4).

Bottom Line: Of note, we found that the majority of CMML patients were classified as World Health Organization CMML-1 and that a 7.5% bone marrow blast cut-point may discriminate prognosis with higher resolution in comparison with the existing 10%.Using random forest survival analysis for variable discovery, we demonstrated that the prognostic power of clinical variables alone is limited.Last, we confirmed the independent prognostic relevance of ASXL1 gene mutations and identified the novel adverse prognostic impact imparted by CBL mutations.

View Article: PubMed Central - PubMed

Affiliation: Department of Malignant Hematology, H Lee Moffitt Cancer Center and Research Institute, Tampa, FL, USA.

ABSTRACT
Since its reclassification as a distinct disease entity, clinical research efforts have attempted to establish baseline characteristics and prognostic scoring systems for chronic myelomonocytic leukemia (CMML). Although existing data for baseline characteristics and CMML prognostication have been robustly developed and externally validated, these results have been limited by the small size of single-institution cohorts. We developed an international CMML data set that included 1832 cases across eight centers to establish the frequency of key clinical characteristics. Of note, we found that the majority of CMML patients were classified as World Health Organization CMML-1 and that a 7.5% bone marrow blast cut-point may discriminate prognosis with higher resolution in comparison with the existing 10%. We additionally interrogated existing CMML prognostic models and found that they are all valid and have comparable performance but are vulnerable to upstaging. Using random forest survival analysis for variable discovery, we demonstrated that the prognostic power of clinical variables alone is limited. Last, we confirmed the independent prognostic relevance of ASXL1 gene mutations and identified the novel adverse prognostic impact imparted by CBL mutations. Our data suggest that combinations of clinical and molecular information may be required to improve the accuracy of current CMML prognostication.

No MeSH data available.


Related in: MedlinePlus