PO:26:096 | From free text to diagnosis: a Natural Language Processing approach for identifying rheumatoid and psoriatic arthritis in the Emergency Department

Antonio Tonutti^1|2, Pierandrea Morandini³, Cosimo Faeti³, Nicoletta Luciano^1|2, Saverio D'Amico³, Antonio Voza^1|4, Victor Savevski³, Carlo Selmi^1|2 | ¹Department of Biomedical Sciences, Humanitas University Pieve Emanuele, Milan, Italy; ²Rheumatology and Clinical Immunology, IRCCS Humanitas Research Hospital Rozzano, Milan, Italy; ³Artificial Intelligence Unit, IRCCS Humanitas Research Hospital Rozzano, Milan, Italy; ⁴Emergency Department, IRCCS Humanitas Research Hospital Rozzano, Milan, Italy

Background. Underdiagnosis and diagnostic delay remain major challenges in rheumatoid arthritis (RA) and psoriatic arthritis (PsA). Emergency Departments (EDs) may be the first point of contact for patients with early symptoms, requiring a high index of suspicion. Natural Language Processing (NLP) offers a promising tool to detect patterns overlooked by clinicians. This study assessed whether NLP applied to ED clinical notes could identify patients who would later be diagnosed with RA or PsA, up to 12 months before their first rheumatology consultation.

Materials and Methods. In this retrospective case-control study, we included patients diagnosed with RA or PsA at our center between Jan 2017 and Dec 2023. For each case, we retrieved ED visits occurring within the 12 months preceding the index rheumatology evaluation. Matched controls (1:10 ratio) were selected among ED patients seen on the same date, with no prior rheumatology evaluation, matched via propensity score on triage “color code.” A supervised machine learning pipeline was applied to ED clinical notes, using pre-trained language models and Bidirectional Encoder Representations from Transformers (BERT), fine-tuned over 5 epochs.

Results. A total of 650 patients accessed the ED in the year preceding RA or PsA diagnosis (58% female; mean age 63±15.6). Of these, 294 (45%) were diagnosed with PsA (7.5% of all new PsA cases diagnosed at our Center; 38% female; mean age 63±14.7), and 356 (55%) with RA (10.4% of all new RA cases; 75% female; mean age 63±16.4). ED visit frequency remained stable in PsA but increased in the three months before RA diagnosis (Figure 1). Musculoskeletal symptoms were the most common presenting complaint. NLP applied blindly to ED notes achieved good discrimination between cases and controls in the training set, but performance dropped on the test set (Figure 2). Word cloud analysis (Figure 3) revealed lexical overlap between true positives and false negatives. Manual review of ED notes was then performed to retain only records referring to suspected arthritis, identifying 151 PsA and 175 RA patients. After reapplying NLP, model performance again showed good discrimination in train set but declined in test set (Figure 4).

Conclusions. This is the first study using NLP to identify RA and PsA from ED notes before specialist referral. A meaningful proportion of patients had ED visits for potential rheumatologic symptoms up to one year prior to diagnosis. These patients were older than general RA/PsA incident cohorts, possibly reflecting greater symptom burden. While NLP demonstrated potential, it failed to consistently distinguish prodromal patients from matched controls, highlighting both the promise and current limitations of this approach in real-world settings.