Sequential drug decision problems in long-term medical conditions: a case Study of Primary Hypertension Eunju Kim ba, ma, msc

8.3.3Data generation

Although the hypertension SDDP model developed in this study focused on the investigation of computational complexity, SDDP modelling is considerably constrained by availability of relevant clinical date. The quality of solution relies on the reliability and robustness of data, which will impact on the accuracy of parameter estimation. Access to relevant patient-level data is necessary to collect the data required for SDDPs, particularly the direct and indirect time-varying treatment effects and the correlation between treatment options. However, these data are unlikely to be collected by conventional RCTs because subsequent treatments are normally not based on random assignment and most interventions are considered as an aggregation of a set of components rather than providing the effects of individual components^[405]. In this respect, the sequential multiple assignment randomised trial (SMART), which involves re-randomisations at each therapeutic stage for identifying optimal sequential treatment strategies in clinical trials, has the potential to provide clinical data regarding sequential decisions^[406]. The primary goal of SMART trials is to identify the best sequencing of treatment options that lead to improved clinical outcomes. SMART trials also provide the effect of certain components of the sequential treatment policy as SMART trials apply factorial designs in a sequential setting. They could also inform time-dependent variables, which are important predictors of the response in the next period and the resulting actions. However, SMART trials are not widely implemented in practice because of feasibility and acceptability^[407]. Executing a SMART trial takes a long-time and costly more than conventional RCTs. It is also uncertain whether the trial design is tolerated by the participants; and whether the assessment procedures and the results are acceptable by clinicians and policy maker. Therefore only a small number of SMART trials have been implemented in “real-world” clinical settings funded by public institutions^{[408, 409]}.

Considering their importance in informing the clinical data and evidence for sequential treatment strategies, potential funding bodies may be interested in whether a SMART trial is feasible and worth pursuing before executing the trial. Modelling an SDDP with existing data may provide preliminary knowledge about the trial design and the likely direction of its effect. The SDDP model could also demonstrate feasibility and the need for SMART trials and thereby assist in their development and implementation. The research questions that can be addressed by a SDDP model may be whether a proposed SMART trial is worth pursuing, and if so what the most feasible design is. Specifically, SDDP modelling could help to address the following questions: how many treatment steps should be allowed, which variables/outcomes should be used to assess treatment response/non-response to inform drug switching, how frequent the patients should be reassessed and how sensitive the outcomes need to be to allow drug switch. The SDDP modelling could also help to decide time and resource implications of the SMART trial, for example, how many participants are required and how long the trial would take.

Four core tasks are proposed for the application of SDDP modelling in preparation for the SMART trial: 1) literature review, 2) modelling of an underlying evaluation model, 3) incorporating an optimisation method into the underlying evaluation model and 4) resulting analysis and discussion.

A comprehensive literature review will be required to understand the key features of the disease and treatment options and to identify the data availability to populate the model. As SDDP modelling covers a broader disease pathway and all possible treatment options, this task could be time-consuming and require the integration of evidence from a range of relevant sources including clinical trials, observational studies, administrative data sets and expert opinion.

The underlying evaluation model is a representation of the real clinical pathway of the disease of interest with drug switching being incorporated. The complexity of the underlying evaluation model depends on the size of the health state space, the assumptions in the transitions between the health states and the complexity of the drug switching rules. IBM such as DES may be more efficient for more effectively describing the complex relationship between the disease pathway and the sequential use of drug(s). However, any decision analytic models such decision-tree and Markov models can be used for the evaluation model if the size and computational complexity of the given problem can be reduced. The evaluation model will estimate the expected total net benefit for different treatment options.

The incorporation of the optimisation method helps to generate the sequential treatment strategies to be evaluated by the evaluation model efficiently. If the SDDP under investigation is relatively simple and manageable, then enumeration can be performed. Otherwise, heuristics (or meta-heuristics) need to be used to identify optimal or ‘near-optimal’ solutions. The applications of these optimisation methods for SDDPs and the general guidance on the trade-offs between key parameters were discussed in this thesis. Their computational efficiency was also demonstrated in the case study of primary hypertension.

Results from the SDDP models can be presented in base-case analysis, scenario analyses and value of information analysis (VOI). The base-case results could be used to rule out sequential drug policies, which are unlikely to be cost-effective, from the proposed SMART trial. Various scenario analyses can help to inform the impact of different trial designs and variable settings on the expected outcome (e.g., time cycle, the follow-up period, sample size, population characteristics and the definition of response/non-response). VOI methods have been proposed as a systematic decision-analytic approach for aiding decision makers in assessing whether there is enough evidence to support new therapies or research for HTA^[410-412]. It is expected that VOI could allow quantifying the potential value of further data collection from SAMRT trials directly from the simulated results. If the VOI exceeds the expected costs of executing the SMART trial, then it is potentially valuable to conduct the SMART trial. Practical issues to conduct the SMART trials, such as time, costs, recruiting participants and staff availability, can be discussed the confirmed design and treatment algorithms of the SMART trial.