Reproducibly Estimating Drug Exposure from Free-text Prescriptions

Background: In pharmacoepidemiology studies, drug exposure is typically estimated from electronic prescribing records. However, the data preparation process is rarely transparent and subject to a range of assumptions. Prescriptions do not always include key information such as start and stop dates, or daily doses, or else are provided as ambiguous free text instructions. The choices and assumptions made when wrangling these data can have profound implications on estimates of drug exposure, resulting in considerable uncertainty in risk attribution of possible adverse events.

Objectives: To develop an open-source software that will advance the efficiency, transparency and reproducibility of drug exposure data preparation.

Methods: We developed a two-stage solution. The first package, doseminer, is designed to extract numeric dosage information from free-text English or Latin dispensing instructions. A second package, drugprepr, then estimates the drug exposure period through a variable combination of defensible data processing decisions.

Results: The two open-source R packages are available for download from https://CRAN.R-project.org/package=doseminer and https://CRAN.R-project.org/package=drugprepr, respectively. The doseminer algorithm represents each free-text prescription instruction using three key metrics: dose number, dose frequency, and dose interval. For each of these three metrics, the range of plausible values is derived. For example, "one or two tablets per day" returns a dose number of 1 (minimum) to 2 (maximum). The drugprepr algorithm is made of eleven distinct preparation steps, covering issues including handling implausible values, dealing with missing data, estimating treatment durations and handling sequential prescriptions with small gaps or overlaps. At each step, the user inputs a decision about how the data will be transformed. By describing the decisions taken at each step, the data preparation process can be replicated. We can vary one or multiple decisions in the data preparation pathway, thereby reporting a range of outcomes and exploring many ‘researcher degrees of freedom'. For example, we have shown the risk of fracture with opioids ranges from HR = 2.64 (95% CI: 2.23, 3.12) to HR = 3.30 (95% CI: 2.78, 3.90) by varying the interpretation of imprecise free text instructions using dose-miner.

Conclusions: The development and sharing of these two R packages has the potential to improve the efficiency, transparency and reproducibility of drug safety research. By sharing the source code, our initial efforts can be further developed by the pharmacoepidemiology community to be used in a wider range of settings and to incorporate further complexities in drug exposure data preparation, for example preparing dosage information.