Abstract
BACKGROUND: Underdiagnosis and delayed diagnosis of rheumatoid (RA) and psoriatic arthritis (PsA) remain major concerns, causing untimely treatments and impacting outcomes. OBJECTIVE: We hypothesize that natural language processing (NLP) may identify red flags from clinical notes in non-specialist settings during early disease phases, thereby improving patient referral. DESIGN AND METHODS: Patients diagnosed with RA or PsA were retrospectively reviewed. Clinical notes from Emergency Department visits occurring in the 12 months prior to the diagnosis were analyzed through NLP into a final classification layer. Propensity score-matched controls without arthritis accessing the same Emergency Department were selected for comparison. RESULTS: Among 650 patients with inflammatory arthritis seeking emergency evaluations in the 12 months before the diagnosis, 294 (45.2%) had PsA and 356 (54.8%) had RA. NLP achieved modest performance in training (AUC(ROC) 70% for RA and 69% for PsA) with a marked drop in independent test sets (AUC(ROC) 62% and 61%, respectively), suggesting limited robustness and accuracy for routine clinical use. Further analysis suggested an overlap between terms included in notes of cases classified correctly versus incorrectly. Manual selection of notes suggestive of arthritis by Rheumatologists did not improve the model performance, which remained of limited generalizability. NLP identified demographic and clinical peculiarities of RA and PsA patients presenting to the Emergency Department, including differing temporal trends of admission between the diseases. CONCLUSION: While NLP showed potential for extracting disease-related signals from Emergency Department clinical notes, its current performance does not support day-to-day clinical implementation. These findings highlight the need for enhanced data quality, interpretability, and algorithm robustness. NLP provided insights into the characteristics of patients seeking emergency care during prodromal phases of RA and PsA.