PharmMT: A Neural Machine Translation Approach to Simplify Prescription Directions

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2785–2796 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 2785 PharmMT: A Neural Machine Translation Approach to Simplify Prescription Directions Jiazhao Li,1 Corey Lester,2 Xinyan Zhao,1 Yuting Ding,2 Yun Jiang,3 V.G.Vinod Vydiswaran4,1 1School of Information; 2Department of Clinical Pharmacy, College of Pharmacy; 3Department of Systems, Populations, and Leadership, School of Nursing; 4Department of Learning Health Sciences, Medical School University of Michigan, Ann Arbor, MI {jiazhaol,lesterca,zhaoxy,dingyt,jiangyu,vgvinodv}@umich.edu Abstract The language used by physicians and health professionals in prescription directions in- cludes medical jargon and implicit directives and causes much confusion among patients. Human intervention to simplify the language at the pharmacies may introduce additional errors that can lead to potentially severe health outcomes. We propose a novel ma- chine translation-based approach, PharmMT, to automatically and reliably simplify pre- scription directions into patient-friendly lan- guage, thereby significantly reducing pharma- cist workload. We evaluate the proposed ap- proach over a dataset consisting of over 530K prescriptions obtained from a large mail-order pharmacy. The end-to-end system achieves a BLEU score of 60.27 against the reference di- rections generated by pharmacists, a 39.6% rel- ative improvement over the rule-based normal- ization. Pharmacists judged 94.3% of the sim- plified directions as usable as-is or with mini- mal changes. This work demonstrates the fea- sibility of a machine translation-based tool for simplifying prescription directions in real-life. 1 Introduction Adverse drug events stemming from medication errors are a vital cause of concern in patient care and are estimated to cost US$42 billion annually or roughly 1% of total global expenditure. In the US alone, medication errors cause one death every day and are responsible for over 700,000 visits to the emergency department and over 100,000 hospitalizations each year (Budnitz et al., 2006, 2011; WHO, 2017). One of the frequent sources of medication er- rors in the US is the directions on the 1.91 billion electronic prescriptions (e-prescriptions) transmit- ted annually (Moniz et al., 2011; Odukoya et al., 2014, 2015). The style and language used in e- prescriptions are highly variable and often filled with medical jargon. For example, in a recent study (Yang et al., 2018), the authors noted that the direction “Take 1 tablet by mouth once daily” was represented in 832 different ways. The study also found that 10.1% of e-prescriptions contained incorrect or confusing language. Pharmacists play a vital role as intermediaries between physicians and patients by translating the rich medical jar- gon in the e-prescriptions written by physicians to patient-comprehensible directions on the prescrip- tion labels printed on pill bottles. However, human translation is time-consuming and subject to errors, potentially leading to medication errors and other patient safety risks due to prescription ambiguity. In this paper, we propose a machine translation- based system, called PharmMT, to simplify the e-prescription directions authored by physicians into patient-friendly language. The system aims to automate the translation and normalization of e-prescription directions and reduce the pharmacists’ overall workload. We investigate multiple neural network-based models, including transformer-based models …

Content truncated to 500 words