OVERVIEW AND EVALUATION OF DECISION-ANALYTIC MODELS FOR THE TREATMENT OF CHRONIC MYELOID LEUKEMIA

Purpose: To describe and analyze the structural and methodological approaches of published decision-analytic models evaluating various treatment strategies in chronic myeloid leukemia (CML) and to derive recommendations for future CML models with a focus on personalized medicine.

Method: We performed a systematic literature search in electronic databases (Medline/PreMedline, EconLit, EMBASE, NHS EED and Tufts CEA Registry) to identify published studies evaluating CML treatment strategies using mathematical decision models. The models were required to compare different treatment strategies and to comprise relevant clinical health outcomes (e.g., responses, progression-free survival, life-years gained or QALYs) over a defined time horizon and population. We used standardized forms for data extraction, description of study design, methodological framework, and data sources. Among other characteristics, we analyzed different modeling types and simulation techniques, endpoints, perspectives, time horizons, uncertainty analysis, model validation and personalized medicine aspects.

Result: We identified 15 different decision-analytic modeling studies. Of these, 14 included economic evaluation. The modeling approaches varied substantially and comprised decision trees, Markov cohort models, state transition models with individual (Monte Carlo) simulation, and mathematical equations. Analytic time horizons ranged from two years to lifetime. Most of the models chose the perspective of the health care system or a societal perspective. Health outcomes included (overall, median, progression-free) survival, life expectancy, and QALYs. Compared treatment strategies comprised bone marrow or peripheral blood stem cell transplantation, conventional chemotherapy, interferon-alpha, and tyrosine kinase inhibitors (TKI). Only one model evaluated a second-generation TKI (dasatinib). The majority of the models did not report a model validation. All models conducted deterministic sensitivity analyses.  In addition, four models reported a probabilistic sensitivity analysis. None of the models evaluated comprehensive personalized medicine strategies.

Conclusion: We found several well-designed models for different CML treatment strategies. However, the quality of reporting varied substantially. We recommend that future models should include novel treatment options such as second-generation TKIs to assess the long-term effectiveness and cost-effectiveness of these treatment strategies, subgroup evaluations for a more personalized decision making, and validation using independent data. Already available models with a short time horizon could be updated with new survival data.