How can we define and analyse drug exposure more precisely to improve the prediction of hospitalizations in longitudinal (claims) data?

Abstract

Background

Risk prediction models can be powerful tools to support clinical decision-making, to help targeting interventions, and, thus, to improve clinical and economic outcomes, provided that model performance is good and sensitivity and specificity are well balanced. Drug utilization as a potential risk factor for unplanned hospitalizations has recently emerged as a meaningful predictor variable in such models. Drug treatment is a rather unstable (i.e. time-dependent) phenomenon and most drug-induced events are concentration-dependent and therefore individual drug exposure will likely modulate the risk. This especially applies to longitudinal monitoring of appropriate drug treatment within claims data as another promising application for prediction models.

Methods and Results

To guide future research towards this direction, we firstly reviewed current risk prediction models for unplanned hospitalizations that explicitly included information on drug utilization and were surprised to find that these models rarely attempted to consider dose and frequent modulators of drug clearance such as interactions with co-medication or co-morbidities. As another example, they often presumed class effects where in fact, differences between active moieties were well established. In addition, the study designs and statistical risk analysis disregarded the fact that medication and risk modulators and, thus, adverse events can vary over time. In a simulation study, we therefore evaluated the potential benefit of time-dependent Cox models over standard binary regression approaches with a fixed follow-up period.

Conclusions

Longitudinal drug information could be utilized much more efficiently both by precisely estimating individual drug exposure and by applying more refined statistical methodology to account for time-dependent drug utilization patterns.