Decision Modeling

Decision models can synthesize evidence from different sources to provide estimates of long-term consequences of a decision with uncertainty. Cohort state-transition models (cSTM) are decision models commonly used in medical decision making because they can simulate hypothetical cohorts' transitions across various health states over time. This tutorial shows how to conceptualize cSTMs in a programming language environment and shows examples of their implementation in R. We illustrate their use in a cost-effectiveness analysis of a treatment using a previously published testbed cSTM. Both time-independent cSTM where transition probabilities are constant over time and time-dependent cSTM where transition probabilities vary over time are represented. For the time-dependent cSTM, we consider transition probabilities dependent on age and state residence. We also illustrate how this setup can facilitate the computation of epidemiological outcomes of interest, such as survival and prevalence. We conclude by demonstrating how to calculate economic outcomes and conducting a cost-effectiveness analysis of a treatment compared to usual care using the testbed model. We provide a link to a public repository with all the R code described in this tutorial that can be used to replicate the example or to be modified to suit different decision modeling needs.

The use of open-source programming languages, such as R, in health decision sciences is growing and has the potential to facilitate model transparency, reproducibility, and shareability. However, realizing this potential can be challenging. Models are complex and primarily built to answer a research question, with model sharing and transparency relegated to being sec- ondary goals. Consequently, code is often neither well documented nor systematically organized in a comprehensible and shareable approach. Moreover, many decision modelers are not formally trained in computer programming and may lack good coding practices, further compounding the problem of model transparency. To address these challenges, we propose a high-level framework for model-based decision and cost-effectiveness analyses (CEA) in R. The proposed framework consists of a conceptual, modular structure and coding recommendations for the implementation of model-based decision analyses in R. This framework defines a set of common decision model elements divided into five components: (1) model inputs, (2) decision model implementation, (3) model calibration, (4) model validation, and (5) analysis. The first four components form the model development phase. The analysis component is the application of the fully developed decision model to answer the policy or the research question of interest, assess decision uncertainty, and/or to determine the value of future research through value of information (VOI) analysis. In this framework, we also make recommendations for good coding practices specific to decision modeling, such as file organization and variable naming conventions. We showcase the framework through a fully functional, testbed decision model, which is hosted on GitHub for free download and easy adaptation to other applications. The use of this framework in decision modeling will improve code readability and model sharing, paving the way to an ideal, open-source world.

Background
People with diabetic retinopathy tend to have lower levels of health-related quality of life than individuals with no retinopathy. Strategies for screening and treatment have been shown to be cost-effective. In order to reduce the bias in cost-effectiveness estimates, systematic reviews of health state utility values (HSUVs) are crucial for health technology assessment and the development of decision analytic models. A review and synthesis of HSUVs for the different stages of disease progression in diabetic retinopathy has not previously been conducted.

Methods/Design
We will conduct a systematic review of the available literature that reports HSUVs for people with diabetic retinopathy, in correspondence with current stage of disease progression and/or visual acuity. We will search Medline, EMBASE, Web of Science, Cost-Effectiveness Analysis Registry, Centre for Reviews and Dissemination Database, and EconLit to identify relevant English-language articles. Data will subsequently be synthesized using linear mixed effects modeling meta-regression. Additionally, reported disease severity classifications will be mapped to a four-level grading scale for diabetic retinopathy.

Discussion
The systematic review and meta-analysis will provide important evidence for future model-based economic evaluations of technologies for diabetic retinopathy. The meta-regression will enable the estimation of utility values at different disease stages for patients with particular characteristics and will also highlight where the design of the study and HSUV instrument have influenced the reported utility values. We believe this protocol to be the first of its kind to be published.