The original project was GM 068968, responding to Joint DMS/NIGMS Initiative to Support Research in Mathematical Biology, PA NSF 02-125. This competing renewal application is continues to propose new mathematical innovations in biomedical computational science and technology. Modeling the pharmacokinetic and pharmacodynamic (PK/PD) behavior of drugs has serious statistical flaws. The PK/PD community still uses mainly parametric methods of modeling based on approximate likelihoods, with no guarantee that studying more subjects will obtain parameter estimates closer to the true values (they often get worse). In contrast, our laboratory has developed methods, both parametric (P) and nonparametric (NP), which are statistically consistent. However, there is still no way to obtain rigorous confidence intervals on P or NP parameter estimates. This is a great weakness. Also, current dosing policies are based only on information available now, though we know we will monitor the patient and adjust dosage in the future. These known future actions are ignored.
Our aims are (1) TO DEVELOP A NEW SEQUENTIAL BAYESIAN METHOD FOR MAKING PK/PD POPULATION MODELS. We propose an exciting new method to obtain rigorous confidence intervals for parameter estimates for both P and NP population PK/PD models. It is an outgrowth of our previous work in GM 068968. It should also provide rigorous confidence intervals on a clinician's ability to hit a desired therapeutic target serum concentration. This will provide a firm mathematical foundation for all population modeling, and for our current work to optimize coordinated combination drug therapy for which we have recently been funded under grant EB 005803. It is also sequential, and thus permits new subjects to be added to a model rather than having to remake it from scratch. This will greatly aid community hospitals to add their own patients to the original model as desired. (2) TO CONTINUE WORK ON OUR ACTIVE CONTROL STRATEGY TO OPTIMIZE LEARNING ABOUT THE PATIENT WHILE TREATING HIM/HER AT THE SAME TIME. Current dosage regimens use only information available up to now. We know we will monitor the patient and adjust dosage in the future. This is ignored. The dosage regimen is not designed to aid in learning about the patient. We now propose to use the dosage regimen as an active partner in the learning process, by calculating how far (and safely) one can deviate a bit from the target goal to probe the patient's system thoughtfully to learn more about it, and thus to maximize therapeutic precision over the projected duration of therapy. We propose to explore future clinical scenarios in advance, now. Our approach is to approximate the Stochastic Dynamic Programming (SDP) equations of Bellman using the IPS (Iteration in Policy Space) algorithm, and a Particle Filter to solve the underlying nonlinear estimation problem. This should make patient care still more intelligent and optimal. ? ? ?
Showing the most recent 10 out of 63 publications
