PK/PD approach to dose selection

Abstract

The utility of a given outcome is the subjective value of that outcome. Decisions differentially affect the probabilities (not the utilities) of possible outcomes. The expected utility of a decision is the average utility of all possible outcomes consequent upon that decision, each one weighted by its probability under the decision. The optimal decision maximizes expected utility. Whilst utilities are subjective, the probabilities of outcomes are objective. Thus, the goal of empirical investigation of dose response should be to provide the particular probability distributions of outcomes as a function of dosage regimen that are required to compute optimal (dosing) decisions. Given the presence of an indication for treatment, and the decision to treat with a given drug, the probability distributions in question are functions not only of dosage (amount and timing), but also of clinical circumstances, i.e., factors that affect individual pharmacokinetics (PK) and/or pharmacodynamics (PD). Let the set of these distributions be called, collectively, the response surface for the drug. Given a mapping from responses to utilities, the response surface is clearly a sufficient condition for optimal dosage decisions. Just as clearly, however, regulatory authorities cannot require a purely empirical estimate of it, as this would entail studying all practically realizable dosage regimens in all possible clinical circumstances, a manifestly impractical task. Perhaps because of the success of empirical hypothesis testing as a means of establishing drug efficacy, that same paradigm has been applied to regulatory requirements for dosing (labeling), with unfortunate results: given that the desired goal is impractical, ad hoc and incomplete strategies have evolved as substitutes. A scientific model-based view, which represents the response surface as parsimonious parametric functions of key PK/PD features, and estimates the surface by pooling data across many different studies despite their different designs resolves the conflict between the empirical demand for data and the decision-theoretic demand for a complete and continuous response surface.