Treatment of chronic diseases such as cancer and depression rely on adaptation of treatments over time to efficiently respond to treatment history and a patient’s evolving health conditions. For example, when treating cancer, physicians prescribe a course of therapy and repeatedly make decisions about the effectiveness of the treatment based on measurable outcomes that include imaging, blood biomarkers, patient-reported symptoms, and other contingencies such as predetermined nonresponse. When one intervention fails or the condition worsens, patients and physicians are consistently faced with the question, “What’s next?"

Despite this need to identify best strategies of care, treatment sequencing is rarely studied in randomized clinical trials. Sequential multiple assignment randomized trials (SMARTs) provide an avenue for testing the comparative effectiveness of patient-centered adaptive treatment strategies (ATSs). SMARTs focus on ATSs that lead to optimal benefit for each individual patient and in turn, for the overall population.

While SMARTs are clinically and methodologically attractive, comparative effectiveness researchers have identified inefficiencies that make the implementation of SMARTs in pragmatic settings challenging. Among these inefficiencies are (1) inflexibility to adjust treatment allocation in future stages according to benefits observed in prior stages, leading to potentially more people being treated with ineffective treatments, (2) unavailability of interim monitoring methods leading to unnecessary continuation of the trial, wasting resources and potentially treating patients further with ineffective ATSs, and (3) inadequate methods to handle missing treatment data and dropouts after baseline, potentially leading to biased and inefficient study results.

The objectives of this project are to:

Aim 1: Develop a pragmatic Bayesian and frequentist adaptive randomization schemes in SMART trials to improve adherence and retention that accounts for cumulatively observed treatment efficacies. The project team will develop these to adequately estimate the ATS effect from such designs and compare them to regular SMART designs via simulations based on various metrics such as bias, efficiency, and average number of patients treated with the most effective ATS.

Aim 2: Develop and evaluate statistically valid interim monitoring approaches to allow for early termination of SMART trials for efficacy and/or futility for binary and survival outcomes. The project team will develop a theoretical framework for interim analyses, resulting in interim monitoring of SMART (IM-SMART) designs and compare them with SMARTs using simulation and reanalyzing existing SMART trials to evaluate the advantages of IM-SMART over SMART in saving resources and avoiding ineffective treatment paths.

Aim 3: Use machine learning to optimize ATSs and handle missing treatment data in SMARTs. Ignoring patients with missing treatment data or dropouts could lead to biased estimates and could also be inefficient due to decreased sample size. Handling missing treatment information in SMART designs is different from the standard methods of handling missing data as the missing treatment information is part of the ATS formulation. The study will use machine learning-based approaches (Q-learning, Super Learners, Ensemble Learning) to develop optimization algorithms for ATSs and handling missing data, evaluate them via simulations, and apply to existing SMART trials.

Aim 4: Develop software packages in various platforms (R, RShiny) and in partnership with a diverse stakeholder advisory board and make them available for comparative effectiveness researcher stakeholders.

The products of this research are valid statistical methods for (1) response-adaptive randomization in SMARTS and corresponding inference, (2) interim analysis of SMARTs, (3) handling missing treatment data in SMARTs, and (4) software packages for the methods developed. These will help our stakeholders (comparative effectiveness researchers) in designing and analyzing patient-centered SMARTs to test individualized treatment strategies to maximize treatment benefit. The project team demonstrates its approaches using three existing SMART trials, one for treating leukemia, another for child neuroblastoma, and the third for treating pain and depression. As SMART designs are now commonly applied in other comparative effectiveness research settings for other conditions, these methods have the highest potential to be disseminated quickly.

The study team includes a statistical methodologist principal investigator, two other statisticians with expertise in SMARTs and optimization machine-learning algorithms, two clinician co-investigators/stakeholders from mental health and oncology, one sleep/mental health statistician/clinical researcher, one patient/statistician stakeholder, and two other statistician stakeholders designing and employing SMARTs in clinical research. All aspects of the project will benefit from feedback and oversight of our Stakeholder Advisory Board.