Improving Outpatient Medication Lists Using Temporal Reasoning and Clinical Texts

An accurate and complete medication list in a patient’s electronic health record (EHR) is critical to prevent medication prescribing and administration errors. Most prior research uses aggregate structured medication data from the EHR to generate and maintain a reconciled list. However, certain critical information for medication reconciliation and decision support exists in free-text clinical notes that may be unavailable in structured data. Structured data in a standard, predictable form can be easily processed by a computer, but narrative data are not codified and thus pose challenges. Natural language processing (NLP) is any system that manipulates free-form text or speech. NLP applications have been developed to identify and extract medical information from non-structured sources; however, few projects have examined the use of NLP as a method for improving the accuracy of medication lists and facilitating medication reconciliation. This study investigated the feasibility of extracting medication information from non-structured electronic clinical sources within the Longitudinal Medical Record (LMR) system, the ambulatory-care EHR at Brigham and Women’s Hospital in Boston. The extracted information can be subsequently used by clinicians at the point of care, thereby reducing prescription and administrative errors.

Specifically, this project: 1) designed and developed a NLP application called Medical Text Extraction, Reasoning and Mapping System (MTERMS), which identifies medication names and drug signatures and other information from free-text clinical notes, encodes medication names using RxNorm and local terminology in the LMR, conducts terminology mapping simultaneously, and structures the extracted information; 2) evaluated the tool by verifying the NLP output against manual review; and 3) identified requirements for a user interface to efficiently use NLP output for medication reconciliation. MTERMs was noted to function with high accuracy, with an F-measure of 90.6 for free-text notes, and 94.0 for structured notes. The F-measure is a measure of a test’s accuracy in which the highest score is 100. For free-text notes, RxNorm covered 98 percent of the terms, while the local drug dictionary only had 83 percent coverage. When mapping between terminologies, only a 62 percent exact match was achieved.

The study showed that the use of NLP to mine free-text medications in order to assemble a list for medication reconciliation is feasible. In addition, such a system could be used to detect potential medication errors. Application to the clinical setting has challenges however, particularly around the updating and mapping of terminologies.