Abstract MOTIVATION: One of the most complex aspects of providing care to cancer patients is building an accurate list of medications a patient is taking. Medication lists need to capture the necessary clinical context about how the patient uses the drug, such as if they can no longer afford a medication or decide to change the frequency at which they take it due to side effects. While hidden from the medication list, these changes are significant to medical decision making. Moreover, many studies focus on providing cancer medication safely. Still, providers are often not aware of other medications, which clinicians may administer in a different practice than the one providing cancer care. While many solutions attempt to improve the medication list, they frequently ignore one of the highest quality sources of medication information: clinical narratives. In this work, we propose utilizing Natural Language Processing (NLP) to extract medications and medication change events from clinical narratives. METHODS & DISCUSSION: In this work, we utilize NLP to extract medications and medication change events from clinical narratives. We use the medication event extraction schema from the N2C2 2022 Challenge to train and evaluate various models for two classification tasks. The first is Event Classification (EC), to identify if a medication was changed. The second is Event Context Classification (ECC) to understand the nature of any change. Using this schema, we can parse a sentence from a clinical note such as “She was experiencing a bad episode of dry cough, so stopped taking lisinopril.” Parsing such a note helps us understand an actor (the patient) initiated an action (stopping the medication) with a temporality (in the past) and with a certainty (not hypothetical). Our approach for both tasks involved using a pre-trained clinical language model while varying the context window around a medication, masking extraneous drugs, and using mark-up tokens around events. The most performant experiments utilized more extended context around a drug, with irrelevant drug masking. We also note that longer-range encoders had better F1 performance than the original BERT architectures. We achieved a 92% performance (micro F1) for EC when detecting a medication change. For ECC, we achieved scores of 74%, 90%, 73%, and 88% (micro F1) over the categories of Action, Certainty, Temporality, and Actor. CONCLUSION: Given the performance observed, this approach could allow clinicians to incorporate additional extracted medication context to improve patient care and provide high-quality treatment plans. Furthermore, we can aggregate this information across patients to give more insights to clinicians and researchers about the actual use of medications. Citation Format: Jacob Hoffman, Neehar Mukne, Daniela Weiss, Christine Swisher, Max Kaufmann. Deep understanding of medication events from clinical narratives is essential to building a holistic picture of cancer patient medication history. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5439.
Read full abstract
Jun 1, 2023
John Whitaker + 7
Open Access
