Optimizing when to start and switch HIV therapy

University Of Pittsburgh At Pittsburgh, Pittsburgh PA

Investigators

Abstract

DESCRIPTION (provided by the applicant): The focus of the proposed research is to employ mathematical techniques to optimize HIV therapy planning while considering patient-specific information such as a tendency to adhere to prescribed drugs. There is much uncertainty about the right time to start patients on antiretroviral therapy or when to switch them to a new therapy. Furthermore, adherence is a major determinant of HIV outcomes and must be considered in the decision-making process. A rigorous, analytical approach to the starting and switching decisions is required to establish best clinical practices. The PI will specifically utilize the well-established methodology of Markov decision processes (MDPs) to derive optimal starting and switching policies under various clinically plausible contexts. MDPs are analytical techniques for solving sequential, stochastic decision problems. This is the typical situation for physicians of HIV patients who may see their patients every 3 months, and need to make a treatment decision without knowing exactly how their patient's health will change over time. Unlike randomized controlled trials, MDPs can often evaluate a large number of potential treatment policies in a short amount of time and without risk to patients. The PI will determine if the resulting policies can be characterized in a general way (such as starting therapy whenever a patient's CD4 count falls below some threshold level). If the number of variables necessary for a complete problem description makes finding optimal solutions computationally intractable, the PI will seek approximate solutions that perform well. In any case, the PI will consult with his committee members to ensure that model development and the resulting solutions are clinically sound. The broader objective of this research is to motivate the further application of MDPs in the medical decision-making community. To date, MDPs have been used to address only a handful of medical problems, even though their modeling and solution framework suggest they can be used more often for such problems.