Background/Purpose: Natural Language Processing (NLP), an interdisciplinary field combining artificial intelligence and language science, has gained significant interest in the medical domain for automated collection and structuring of medical data. In this study, we employed an NLP approach to identify and extract the characteristics of patients with D2T RA from their computerized medical records.

Methods: We conducted a monocentric observational retrospective cohort study in our French hospital. Patients were longitudinally recruited over a 5-year period (2015-2020), and their characteristics were retrieved from their electronic health records spanning 2010-2022. Firstly, we performed an NLP phenotyping tool to identify patients who had at least one hospital stay related to RA between 2015 and 2020. The tool relied on CIM-10 coding and the recognition of specific keywords (e.g., “rheumatoid arthritis” or “ACPA”) in the medical records to calculate the patient’s probability of meeting the RA phenotype. We subsequently employed a named-entity recognition algorithm to identify all drugs associated with RA from the patients’ medical reports. The drugs were then arranged in chronological order based on their appearance, and the date of introduction was approximated as the date of first extraction. D2T RA was defined as the failure of at least two targeted therapies (identified by the algorithm retrieving three or more biologics), thereby allowing us to separate our RA cohort into D2T and non-D2T RA groups. Finally, we used another NLP program employing named-entity recognition to retrieve the characteristics of these patients. Specific keywords related to general variables (sex, age at first stay), disease characteristics (presence of joint erosion, DAS28 activity score), and comorbidities were sought in the medical reports using cumulative extraction. The occurrence of these specific tokens was linked to the presence of the respective variable. Finally, a multivariate logistic regression analysis was performed to assess the relationship between variables and the D2T group.

Results: The phenotyping tool identified 972 RA patients, of which 313 were classified as D2T RA and 659 as non-D2T RA. The presence of joint erosions (OR 2.97 CI95% [2.17; 4.08], p< 0.001) and higher disease activity (median DAS28 3.16 vs 2.89) were associated with the D2T group. Female gender, obesity, smoking, anxio-depression, and fibromyalgia comorbidities showed similar proportions in our two groups. Chronic kidney disease (OR=0.66 CI95% [0.44;1.01], p=0.068), cardiac insufficiency (OR 0.52 CI95% [0.38;0.7], p=0.001), liver insufficiency (OR=0.7 CI95% [0.35;1.4], p=0.39), and a history of malignancy (OR=0.62 CI95% [0.47;0.82], p=0.001) were found in higher proportions in the non-D2T group.

Conclusion: NLP facilitates the automated collection of diverse data directly from medical reports using named-entity recognition algorithms. In this study, we demonstrated its application in identifying potential contributing factors to D2T RA. Enhanced understanding of the mechanisms underlying D2T RA and early detection of contributing factors hold promise for improved outcomes and management in this heterogeneous population.

« Back to ACR Convergence 2023

ACR Meeting Abstracts - https://acrabstracts.org/abstract/identification-of-contributing-factors-to-difficult-to-treat-rheumatoid-arthritis-d2t-ra-in-a-cohort-of-972-patients-using-a-natural-language-processing-nlp-approach/