A data-driven approach to optimized medication dosing: a focus on heparin.

Abstract

To demonstrate a novel method that utilizes retrospective data to develop statistically optimal dosing strategies for medications with sensitive therapeutic windows. We illustrate our approach on intravenous unfractionated heparin, a medication which typically considers only patient weight and is frequently misdosed. We identified available clinical features which impact patient response to heparin and extracted 1,511 patients from the multi-parameter intelligent monitoring in intensive care II database which met our inclusion criteria. These were used to develop two multivariate logistic regressions, modeling sub- and supra-therapeutic activated partial thromboplastin time (aPTT) as a function of clinical features. We combined information from these models to estimate an initial heparin dose that would, on a per-patient basis, maximize the probability of a therapeutic aPTT within 4-8 h of the initial infusion. We tested our model's ability to classifying therapeutic outcomes on a withheld dataset and compared performance to a weight-alone alternative using volume under surface (VUS) (a multiclass version of AUC). We observed statistically significant associations between sub- and supra-therapeutic aPTT, race, ICU type, gender, heparin dose, age and Sequential Organ Failure Assessment scores with mean validation AUC of 0.78 and 0.79 respectively. Our final model improved outcome classification over the weight-alone alternative, with VUS values of 0.48 vs. 0.42. This work represents an important step in the secondary use of health data in developing models to optimize drug dosing. The next step would be evaluating whether this approach indeed achieves target aPTT more reliably than the current weight-based heparin dosing in a randomized controlled trial.