statistical modeling of calgb 80405 (alliance) identifies ... · male (blue) strata. dot size and...
TRANSCRIPT
-
Results
Methods
• CALGB 80405 is a Phase III clinical trial of mCRC patientsthat evaluated first line FOLFOX or FOLFIRI in combinationwith randomly assigned cetuximab or bevacizumab. Primaryresults did not show a statistically significant difference inarms.• Sidedness has emerged as an important prognostic factorin treatment decisions for mCRC patients.• However, side-dependent risk factors are currently notknown. These factors may be important for further refiningmCRC prognosis and informing treatment decisions.• In order to examine the role of right versus left sideness inmCRC, we built multivariate predictive models utilizing anovel, hypothesis-free machine learning approaches.
• We used 99 baseline and demographic variables from theCALGB 80405 study population (n=2,334). 1,904 patientshad information about left versus right tumor side and wereincluded in our analysis. Of these, 949 were KRAS wild-type(primary cohort).• The primary endpoints were overall survival (OS) andprogression-free survival (PFS).• To individually model all combinations of 99 potentiallypredictive variables (i.e 99!) is prohibitively complex and timeconsuming. Therefore, we used a Monte Carlo BayesianGeneralized Linear Model analytical platform, REFSTM, toidentify a set of highly optimized and parsimoniousmultivariate models (ensemble) most likely to predict OS andPFS. This approach minimizes the risk of overfitting.
• By evaluating the importance of a particular variables withineach model in the ensemble, we estimated the uncertaintysurrounding each variable and, through ensemble frequency,we identified key variables that are consistently important.• Sub-group models for side and for gender examineddifferences in drivers of disease based on this stratification.
Statistical modeling of CALGB 80405 (Alliance) identifies influential factors in metastatic colorectal cancer (mCRC) dependent on primary (1o) tumor side
•Leon Furchtgott1*, David Swanson1*, Boris Hayete1, Iya Khalil1, Diane Wuest1, Kelly Rich1, Robert Miller1, Andrew B. Nixon2, Donna Niedzwiecki3, Jeffrey A. Meyerhardt4, Eileen Mary O'Reilly5, Fang-Shu Ou6, Heinz-Josef Lenz7, Federico Innocenti8, Alan P. Venook9,
Alliance for Clinical Trials in Oncology•1GNS Healthcare, Cambridge, MA; 2Duke University Medical Center, Durham, NC; 3Duke University, Durham, NC; 4Dana-Farber Cancer Institute/Partners CancerCare, Boston,
MA; 5Memorial Sloan-Kettering Cancer Center, New York, NY; 6Mayo Clinic Cancer Center, Rochester, MN; 7University of Southern California Norris Comprehensive Cancer Center, Los Angeles, CA; 8University of North Carolina, Chapel Hill, NC; 9University of California, San Francisco, San Francisco, CA
• Primary side, AST level, ECOG PS, and siteof metastases play a central role inexplaining variation in OS and PFS.• Depending on the primary tumor side,different factors appear to drive OS:
• Left sided: urine protein level andtreatment intent for left-side mCRC• Right sided: liver and lung site ofdisease for right-side mCRC.
• Different factors may impact OS for menand women.• Our findings suggest that side- and gender-specific variables may be important forpredicting mCRC course and survival.• Additional research, including prospectivestudies and evaluation of biologic pathways,is necessary to confirm these findings.
Support: U10CA180821, U10CA180882, EliLilly and Company, Genentech, Pfizer.
Background
Conclusions
Outcome (OS or PFS)
Demographic variable(s)Lab, pathology, imaging etc. variable(s)Medication variables, including chemotherapyPast medical history
Figure 1: Schematic of REFS multivariate predictive inference platform
Figure 3: Forest plot showing key variables predictive of PFS in full cohort. Dot size and shading proportional to ensemble frequency.
Figure 2: Forest plot showing key variables predictive of OS in primary cohort. Dot size and shading proportional to ensemble frequency.
Progression-free survival (full cohort): Key variables predictive of PFS include AST concentration, ECOG performance status, use of full-dose coagulants, and lung and abdominal site of disease indicators.
Figure 4: Forest plot showing key variables predictive of OS in left (green) and right (purple) strata. Dot size and shading proportional to ensemble frequency.
Primary side stratum-specific models (OS): In primary side stratum-specific models, urine protein level, treatment intent (palliative as reference)and hemoglobin concentration were more associated with left-side survival, while liver and lung sites of disease were more associated with right-side survival.
Figure 5:Urine protein level is associated with left-side survival (left Kaplan-Meier plot) but not right-side survival (right plot)
Figure 6:Treatment intent is associated with left-side survival (left Kaplan-Meier plot) but not right-side survival (right plot)
Gender stratum-specific models (OS): In gender stratum-specific models, women were moreinfluenced by metastatic status, BMI, and liver site of disease; men were more influenced by urineprotein level and diabetes status.
Figure 7: Forest plot showing key variables predictive of OS in female (red) and male (blue) strata. Dot size and shading proportional to ensemble frequency.
Figure 8: Diabetes is associated with survival among men (bottom Kaplan-Meier plot) but not women (top plot)
Model 1
Model 128
Model exploration for single model
Ensemble of 128 optimized models
Overall survival (full cohort): Key variables predictive of OS include primary side, AST concentration, ECOG performance status, treatment intent (palliative reference), and local primary and abdominal site of disease indicators.