(193) Medication Data in Research: Improving Completeness Using Rule-based Algorithms on Textual Variables

Background: Generation of structured dose instructions is pivotal in pharmacoepidemiology studies, allowing for precise drug exposure calculation. Electronic medical records (EMR) contain medication prescribing data that include both discrete and textual order information. However, discrete data are often incomplete, and textual data fields are often unusable. Manual extraction of discrete data from textual fields can be impractical in large datasets and are prone to missingness.

Objectives: Our work aims to efficiently structure textual data into more complete and usable variables using algorithmic rules, enhancing the reliability of the research.

Methods: We accessed two datasets from the Indiana Network for Patient Care as part of an ongoing clinical trial. The datasets included outpatient prescribing information for opioids (n=1523 orders) and buprenorphine (n=431) for 267 patients. We developed rule-based algorithms in SAS 9.4 that scan unstructured text for keywords using indexing functions and matching to fill missing values or create new variables. Essential information such as strength, dispensed quantity, duration, frequency, route, and units per administration were considered. Prescribed daily dose (PDD) (product of strength and quantity ordered divided by duration) was computed both before and after the application of the algorithm. The final algorithm was applied to both datasets and used a zero-assumption approach.

Results: Among buprenorphine orders, application of the algorithm resulted in a decrease of missingness from 93.5% to 0.7% for strength, 8.4% to 3.5% for frequency, 28.8% to 0% for units per administration, and 34.1% to 10.4% for duration. Similarly, in the opioid order dataset, missing values were reduced from 72.4% to 0.9% for strength, 7.0% to 1.1% for frequency, 13.8% to 3.0% for units per administration, 3.2% to 3.1% for quantity ordered, and from 4.9% to 3.3% for duration. PDD could be calculated in 3.9% of all orders in the buprenorphine data before application of the algorithm and 89% after. Among all opioid orders the PDD could be calculated in 14.7% of orders before the algorithm and in 96% after application of the algorithm. Algorithm decisions were reviewed manually by clinicians to confirm validity.

Conclusions: Substantial reductions in missing values were observed after applying the algorithm to the data resulting in higher ability to generate measures of medication exposure. This work demonstrates the potential of text mining in reducing missingness of prescribed medication data and improving its usability for further analysis. This work demonstrates the value of techniques to further enhance the completeness of EMR data.