A Cascaded Regression Approach for Precision Medication Dosing
We present a cascaded generalized linear modeling approach, which generates a personalized medication dosing policy for patients in critical care environments. We validated our approach using retrospective data from 4,470 patients extracted from a publicly available clinical data archive. We have found our approach to be nearly twice as effective in classification of patient state (under-dosed, over-dosed, or therapeutic) when compared to the performance of clinicians. This paper is an important illustration of how to leverage large-scale retrospective clinical databases to develop precision clinical policies. Our approach is particularly valuable in the case of medication dosing, where clinical trials are unable to dose highly ill patients, or provides doses at levels that may be deemed "unsafe". Hence, retrospective analysis, like ours, allows us to more precisely characterize this personalized dose-response relationship.