Handling time-varying confounding in state transition models for dynamic optimization of adaptive interdisciplinary pain management

Interdisciplinary pain management combines multiple disciplines of professionals to understand the biological and psychosocial factors causing a patient's pain and to determine the best treatments among many to administer. To improve current and future pain outcomes, the developed adaptive interdisciplinary pain management framework employs approximate dynamic programming with state transition and outcome models estimated from actual patient data. The sequential treatment structure and observational nature of the data lead to a form of endogeneity, which results in biased model parameter estimates when developing state transition and outcome models. This research develops a process based on the inverse probability of treatment weighted method to address the endogeneity in estimating state transition and outcome models. This article discusses a general method developed for independent treatments. The proposed approach can potentially be employed for adaptive treatment in other sequential health care applications based on observational data. Our approach is demonstrated using data from the Eugene McDermott Center for Pain Management at the University of Texas Southwestern Medical Center at Dallas.