Targeted sampling to observe individual patient pharmacokinetics is an important tool for precision dosing of antimicrobials. In conjunction with an applicable statistical model, even a small number of samples can be used to generate individual-specific estimations to guide dosing decisions. This individual-specific inference has been widely conducted by maximum-a-posteriori (MAP) estimation. This procedure is a point estimator, generating a single ‘best’ value for each of the parameters. Using the MAP estimate to make a decision neglects uncertainty; while this is in some sense the ’best’ candidate, it is not known if it is meaningfully better than any other value. The individual uncertainty cannot be meaningfully inferred from the collective behaviour of point estimates.
Technical barriers hinder uncertainty quantification for these individual estimates in classical statistics. In a fully Bayesian approach, the entire joint posterior distribution of the individual parameters is directly obtainable. Using simulations based on vancomycin pharmacokinetics, it is demonstrated that popular strategies based on peak and trough sampling are poorly informative. A statistical approach is outlined to formally evaluate the performance of the process, in both an absolute sense, and relative to prior model predictions. Performance was sensitive to characteristics of the simulated population, especially its residual error and between-subject variation, but all sparse sampling strategies were sufficiently uncertain to undermine practical applications, and the gain of information relative to population predictions from the base model was modest. These observations demonstrate an urgent need for contextual uncertainty quantification in model-informed precision dosing, and illustrate a candidate workflow to achieve it.