(277) "Take Up to Eight Tablets Per Day": The Impact of Incorporating the Uncertainty of Free-Text Medication Instructions for Drug Exposure Data Preparation

Publication Number: 608

Background: Prescription data from routinely collected health data often provides drug exposure information for pharmacoepidemiology. Start and stop dates and dosage are derived from prescribing information, both structured data and unstructured free-text instructions such as “take two tablets per day” or the more imprecise “take up to eight per day”. Data preparation of drug exposure information is rarely fully reported, yet assumptions made during this stage can have considerable implications for pharmacoepidemiology research. This step may have bigger consequences for “pro re nata” (PRN) drugs.

Objectives: Using the worked example of opioids and fracture risk, to examine the impact of incorporating narrative prescribing instructions and subsequent drug data preparation assumptions on adverse event rates.

Methods: R-packages for extracting free-text medication prescription instructions in a structured form (doseminer) and an algorithm for transparently processing this information (drugprepr) were developed (available on CRAN). Data from Clinical Practice Research Datalink were used to identify adult patients without cancer, with at least 1 opioid prescription during 2017. To define drug exposure, prescription duration was defined as (quantity of tablets x drug interval)/ numerical daily dose [ndd]). Ndd was defined as drug number x drug frequency. Doseminer was used to extract numerical values from free text (e.g. “take up to eight tablets/day” to dose number 0-8; dose interval 1 day). Drugprepr allowed computation of ndd using different minimum, mean, maximum drug number and frequency assumptions. We tested the impact of varying decisions by estimating the risk of opioids on fracture risk using Cox proportional hazards models compared to unexposed person time. Fractures were attributed to opioids if the patient was on therapy at the time of the event.

Results: During the study window, 60,394 patients were identified with 190,754 opioids prescriptions of which 41.6% prescriptions were PRN. Following application of the two R-packages, variation in the decisions regarding ndd resulted in marked differences in drug exposure impacting person years of exposure (5,439-10,190 pys) and the number of events attributed to the drug (n=301- 422). The distribution of hazard ratio as a function of the decisions ranged from 2.64 (95% CI: 2.23, 3.12) to 3.30 (95% CI: 2.78, 3.90).

Conclusions: Assumptions made during the drug preparation process, especially for PRN drugs or those that have variability in prescription instructions, can impact the results of subsequent risk estimates. Applications of the above packages can improve transparency related to drug preparation assumptions, in line with best practice advocated by international pharmacoepidemiology guidelines.