03/14/2014 10:54 am ET Updated May 14, 2014

Being SMART about Constructing Dynamic Treatment Regimes

By Susan Murphy and Inbal Nahum-Shani

Dynamic treatment regimes (DTRs) are of growing interest across the medical field as these regimes provide a way to operationalize decision making concerning sequences of individualized treatments.

A dynamic treatment regime is composed of a sequence of decision rules. Each decision rule maps up-to-date patient information to a recommended treatment. The patient information includes not only patient characteristics but also intermediate outcomes collected during the intervention, such as the response, adherence and engagement in treatment. The decision rules detail how treatment components (e.g., the type of treatment, the intensity/dosage or the timing of treatment provision) are adapted over time in response to the unique and changing needs of patients as they progress in treatment.

DTRs also are known as adaptive interventions, adaptive treatment strategies, multi-stage treatment strategies, and treatment policies. In the field of statistics, P.W. Lavori, J.M. Robins and P.F. Thall were among the first scholars to investigate and contribute to this area.

From a clinical standpoint, there are many reasons to consider DTRs, particularly in the context of preventing and treating chronic disorders such as obesity, depression, alcohol use and other addictions.

First, the same treatment might not be sufficiently beneficial for most or all patients. Hence, caregivers may have to try a series of treatments in order to achieve sufficient response.

Second, the effectiveness of the treatment might change over time due to the waxing and waning course of the disorder, or due to changes in risk and resiliency. Hence, it is important to determine if and when the treatment is no longer working for a specific patient and, accordingly, consider subsequent treatment options or tactics (e.g., switching to another type of treatment or enhancing the dose of the current treatment).

Third, comorbidities are common in the context of many chronic disorders, for example, comorbid HIV infection and substance use and comorbid depression and alcoholism. In the treatment of comorbid disorders, it may be important to treat the disorders sequentially so as to minimize treatment burden.

Fourth, relapse is common among many individuals struggling with chronic disorders. Hence, maintenance treatments should be considered after an acute response so as to prevent relapse and address relapse when and if this occurs.

Fifth, intensive, longer-term, interventions often involve greater cost, burden and/or side effects. This motivates the development of interventions in which the intensity of the treatment is reduced when possible.

Finally, the related difficulty of maintaining treatment adherence highlights the need to construct interventions in which the type or intensity of the treatment is tailored to encourage adherence and engagement in treatment. Overall the DTR approach has the potential to facilitate the development of more efficacious interventions and to improve the management of many chronic diseases.

The Sequential Multiple Assignment Randomized Trial (SMART) is an experimental design that was developed specifically to aid in the construction of efficacious DTRs. A SMART involves multiple randomizations that are sequenced over time. Each randomization corresponds to a critical decision or a critical open question concerning the sequencing and adaptation of treatment options.

Increased interest in the construction of DTRs has led to the conduct of a number of SMART studies in many fields, including obesity, substance abuse disorders, and mental health disorders. A partial list of SMART studies that are completed or currently in the field can be found here.

A variety of data analysis methods can be used with SMART to inform the construction of efficacious DTRs. These include methods for comparing DTRs that are embedded in a SMART design (e.g., the Weight and Replicate); as well as methods that enable the investigation of candidate tailoring variables that are not integrated in the SMART study (e.g., Q-learning).

This field generates many research opportunities for statisticians, including the development of improved experimental designs and how best to size these study designs along with analytical methods that can aid clinical scientists to construct high-quality (i.e., efficacious) DTRs. There are many exciting methodological issues relating to the analysis of the study data including how best to construct an optimal sequence of decision rules, how to construct confidence intervals for parameters in the decision rules, how to address problems with non-compliance and study attrition, the appropriate use of surrogate outcomes in constructing the decision rules and how to conduct mediational analyses.

An exciting extension of DTRs are Just-In-Time Adaptive Interventions (JITAIs), which are now possible due to advances in mobile and wireless technology. Similar to a DTR, a JITAI is operationalized via decision rules. However, in a JITAI, the tailoring variables can be assessed any time during the day-to-day lives of patients via passive (e.g., wireless sensors) or active (i.e., self-report) data collection modes. Moreover, the treatment options in a JITAI can be delivered in real-time, and are typically in the form of feedback, behavioral and cognitive strategies, information and prompts that support the ability of the patient to self-manage his/her chronic condition.

Given the frequent and real-time adaptation of treatments in JITAI, further advances in experimental designs and analytical methods are needed in order to inform the sequencing and adaptation of treatment options in a JITAI.

The mission of the Statistical Reinforcement Learning Lab at the University of Michigan is to develop a suite of innovative learning algorithms, including treatment designs, experimental designs and data analysis methods to help investigators obtain and utilize empirical evidence necessary for developing high-quality JITAIs. Click here for more information.

Susan Murphy is the H.E. Robbins Professor of Statistics and a Research Professor at the Institute for Social Research at the University of Michigan.

Inbal (Billie) Nahum-Shani is a Research Assistant Professor at the Institute for Social Research at the University of Michigan.