Background: To dissect the most efficient method for prioritization is a desirable goal in appointment triaging of new patients in rheumatology departments. Digital aids, partly augmented by artificial intelligence (AI) have found their way into real-world practice and have been evaluated for diagnosis and monitoring.

Objectives: Early presentation to a rheumatologist, e.g. within six weeks in case of active arthritis, is a justified request, that cannot always be met due to scarce resources.

Methods: We analyzed parameters of interest (reason for referral including clinical complaints and laboratory findings) obtained from initial requests by referring practitioners to a single tertiary rheumatology center in Germany. Upon review of the submitted information, an experienced (>3 years) rheumatologist attributed a time frame for the appointment (≤6 weeks, ≤3 or >3 months) depending on his assessment of urgency and probability for an inflammatory rheumatic disease (IRD). The waiting time between initial request and the actually realized appointment were compared between IRD and non-IRDs. A decision tree (DT), derived from the field of supervised learning within AI was established, using all available parameters from the initial request as input. To compare waiting times, a theoretical appointment was assumed after 60 days for suspected IRD and after 120 days if an IRD was deemed less likely. After that, accuracy and simulated waiting time were calculated and compared.

Results: A total of 800 initial presentations (median age 53 years (IQR: 39-63), including 69.4% females) were analyzed between 2021 and 2023. An IRD was confirmed in 409 cases (51.1%), 193 of which with rheumatoid arthritis ( Table 1 ) and were attributed to a waiting time of 58 days (vs. 93 days in non-IRD patients, p<0.01, Table 2 ). Submitted information was incomplete in many cases. Overall, C-reactive protein (CRP) was reported in about 50%, erythrocyte sedimentation rate (ESR) in 31%, status of the rheumatoid factor (RF) in 27%, anti-CCP in 34%, antinuclear autoantibodies (ANA) in 17% and HLA-B27 in 4%. Regarding symptoms, 65% explicitly reported arthralgia/pain, 17% joint swellings and 16% joint stiffness. Stratification using the DT, incorporating all available parameters, yielded an accuracy of 67% and predicted a 19% reduction in waiting time. Notably, accuracy improved to 78% when the analysis was restricted to cases with known basic laboratory results. Stratification by use of laboratory findings only (if either CRP, RF, anti-CCP, or HLA B27 found positive, qualified for earlier appointment) resulted in a notably less accuracy (62%) and potential time savings (9%). Over all instruments, sensitivity was found considerably inferior to specificity.

Conclusion: Manually reviewed information submitted for rheumatology referral though time- and resource-consuming performed well in stratification. Complex AI-supported models were found useful, especially if laboratory results were complete and lower sensitivity is acceptable.

REFERENCES: NIL.

Acknowledgements: NIL.

Disclosure of Interests: None declared.