OBJECTIVES: Multi-criteria decision-analysis (MCDA) is an important tool for assessing the benefits and risks of medical products. Existing probabilistic MCDA methods often require: dichotomization of time-to-event (e.g., overall survival, OS) and ordinal outcomes (e.g., RECIST); and the routinely unrealistic assumption that utility functions follow a linear relationship. Our new framework, Bayesian Multi-Criteria Augmented Decision Analysis (MCADA), accommodates time-to-event and ordinal outcomes without dichotomization and non-linear utility aggregation. Here, we apply MCADA to two randomized clinical trials (RCTs) in oncology.

METHODS: Our analyses considered OS, RECIST, and grade 3/4 adverse events (AEs) from two phase II RCTs (NCT01439568 and NCT01124786) that had inconclusive primary analyses. We used patient-level data to calculate utility scores by study arm for pairwise benefit-risk assessment. We used linear and non-linear (Emax and logistic) utility aggregation functions. Weights were determined using analytical hierarchy process: OS weighted twice as important as RECIST, and RECIST twice as important as AEs (utility weights of 0.57, 0.29, and 0.14, respectively).

RESULTS: In the first case study on small-cell lunger cancer, we found the treatment (LY2510924) added to standard-of-care (carboplatin + etoposide) had lower utility compared to the standard-of-care alone under the linear and non-linear utility assumptions. Under the linear utility function, the mean difference in utility was -0.09 (95% credible interval [95%CrI]: -0.22, 0.02). The second case study on metastatic pancreatic cancer showed lower utility for experimental treatment (CO-1.01) compared to the standard-of-care (gemcitabine; mean: -0.06; 95%CrI: -0.15, 0.04).

CONCLUSIONS: In our case studies, MCADA was able to point to lack of clinical benefit in experimental therapies more clearly. MCADA can be used as a supplemental analysis to facilitate go/no-go decisions in early clinical development programs for oncology. MCADA can accommodate variables in their natural forms without imposing linear utility assumptions and be used to as an optimal tool for clinical development.