Summer 2007 Program on Challenges in Dynamic Treatment Regimes and Multistage Decision-Making
The management of chronic disorders, such as mental illness, substance dependence, cancer, and HIV infection, presents considerable challenges. In particular the heterogeneity in response, the potential for relapse, burdensome treatments, and problems with adherence demand that treatment of these disorders involve a series of clinical decisions made over time. Decisions need to be made about when to change treatment dose or type and regarding which treatment should be used next. Indeed, clinicians routinely and freely tailor treatment to the characteristics of the individual patient with a goal of maximizing favorable outcomes for that patient. To a large extent the tailoring of sequences of treatments is based on clinical judgment and instinct rather than a formal, evidence-based process.
These realities have led to great interest in the development of so-called "dynamic treatment regimes" or "adaptive treatment strategies." A dynamic treatment regime is an explicit, operationalized series of decision rules specifying how treatment level and type should vary over time. The rule at each stage uses time-varying measurements of response, adherence, and other patient characteristics up to that point to determine the next treatment level and type to be administered, thereby tailoring treatment decisions to the patient. The objective in developing such multistage decision-making strategies is to improve patient outcomes over time.
Methodology for designing dynamic treatment regimes is an emerging area that presents challenges in two areas. First, experimental designs for collecting suitable data that can be used efficiently to develop dynamic regimes are required. Second, techniques for using these and other data to deduce the decision-making rules involved in a dynamic regime must be developed. In both areas, input from researchers in a variety of disciplines and collaborations among them will be critical.
Trials in which patients are randomized to different treatment options at each decision point have been proposed; however, little is known about when such trials should be conducted in lieu of the current approach of melding clinical judgment and expert opinion to formulate decision rules and using the standard two-group paradigm. An alternative approach is to conduct a series of randomized trials, as in agriculture and engineering; again, there is little guidance on how to implement this approach when the goal is to develop a dynamic regime.
Methods to make use of data in developing dynamic regimes involve complex considerations. The construction of optimized decision rules requires incorporating the effects of future decisions when evaluating present decisions, as is well-known to scientists working on improving multistage decision-making. Treatment given at any time may set a patient up for improved response to subsequent treatments or have delayed effects that either enhance or reduce effectiveness of subsequent treatments. The development of a dynamic regime hinges on how one operationalizes the relative importance of patient outcomes over time. Researchers who work on multistage decision problems in other contexts (robotics, artificial intelligence, control theory) readily recognize these types of issues. A key challenge is to determine how to collect sufficient information to ascertain the "state" of an individual insofar as making treatment decisions goes. Typically, a great deal of information is available at each decision point, and methods for feature extraction developed by statisticians and computer scientists are well-suited to this problem, but the focus on multistage decision-making rather than prediction requires evaluation of these methods from a different perspective.
Computational and inferential challenges arise in all of these endeavors; e.g., complexities of optimizing dynamic regimes can invalidate standard statistical inferential techniques, scientific considerations entail thinking beyond the standard loss functions familiar to statisticians, and the abundance of information at each decision point quickly leads to a "small n, large p" problem and the attendant computational issues. For some disorders, e.g., HIV infection, knowledge of the underlying within-subject biological has led to development of sophisticated mechanistic models for the processes governing disease progression and effect of treatment, which offer a scientific basis (via closed loop control methods) for designing dynamic regimes; however, this approach has not been widely explored or tested in samples of patients in this context.
This SAMSI summer program brought this area to the attention of statistical and applied mathematical scientists, whose expertise is critical; jump-start the necessary methodological development; and nurture the necessary interdisciplinary collaboration and communication between statisticians/applied mathematicians and computer scientists and health and behavioral science researchers.
Program Leaders: Susan Murphy (University of Michigan), Daniel Scharfstein (Johns Hopkins Bloomberg School of Public Health), Joelle Pineau (McGill University); Local Scientific Coordinators: Marie Davidian and Butch Tsiatis (North Carolina State University).