740. Does Natural Language Processing Improve Identification of Hospitalized People Who Use Drugs?. (15th December 2022)

• Record Type: Journal Article
• Authors: Sato, Taisuke; Murphy, Meghan; Miller, Robert; Wurcel, Alysse Gail
• Is Part Of: Open forum infectious diseases. Volume 9:(2022)Supplement 2
• Publication Date: 2022-12-15
• Subjects: Communicable diseases -- Periodicals; Medical microbiology -- Periodicals; Infection -- Periodicals; 616.9
• DOI: 10.1093/ofid/ofac492.031
• Languages: English
• ISSNs: 2328-8957

Abstract

Abstract: Background: In the absence of substance use disorder treatment and harm reduction tools, people who use drugs (PWUD) are at increased risk for severe injection-related infections (SIRI). Health services research characterizing the scope and trends of hospitalizations for SIRI often requires administrative billing records, urine toxicology, and medications to treat opioid use disorder, often supplemented with manual chart review. The goal of this research was to examine the feasibility of using natural language processing (NLP) to augment traditional systems of PWUD. We explored the practicality of using Natural Language Processing as a tool to accurately identify PWUD in electronic medical records (EMR). Methods: We worked with information technology/computer science specialists to create a cohort of all people hospitalized at Tufts Medical Center from January 1, 2020, to April 1, 2022. Patients entered this cohort through the presence of any of these criteria: ICD10 codes for substance use disorder, positive urine toxicology results for drugs (including opiates, cocaine, amphetamine), prescriptions for medications to treat substance use disorder (at admission, during hospitalization or discharge) and the presence of NLP keywords. The NLP keyword list was created through iterative meetings with IT specialists (Table 1). Results: There were 3102 patients with 3501 hospitalizations. The Venn Diagram (Figure 1) displays the overlap of the 4 criteria. Only 55 people were identified by NLP terms only. The most frequent criteria filled for cohort entry was the toxicology screen. Only 4.1% of the group had all 4 criteria. Figure 1: Venn diagram, SUD Cohort Breakdown by PWUD Identification Methods Conclusion: Identifying PWUD using an NLP algorithm was possible, but there was a small number of people who were added to the cohort only on NLP criteria. Improved methodologies of identifying PWUD with medical needs are necessary to better identify gaps in healthcare, facilitate benchmarking of quality care, and improve the lives of PWUD. Disclosures: All Authors : No reported disclosures.