Background/Purpose: To determine if a computer-based triage tool can accurately classify referrals as inflammatory or non-inflammatory using information obtained from the patient; not requiring assessment by healthcare workers.

Methods: Patient referrals to a single rheumatologist were studied. Patient lists for 11 commonly encountered diagnoses [rheumatoid arthritis (RA), psoriatic arthritis, lupus, osteoarthritis, gout, soft tissue rheumatism (i.e. arthralgia NYD, tendonitis, mechanical low back pain), ankylosing spondylitis (AS), polymyalgia rheumatica (PMR), and fibromyalgia] were created. The diagnoses of RA, lupus, AS, PMR, and other seronegative spondyloarthropathies were classified as inflammatory arthritis (non-crystalline) (IA). At least five patient charts for each diagnosis were reviewed and eligible for study if a self-reported patient pain diagram was also completed. The following patient-reported information was collected from each referral: age, sex, symptom duration, and pain diagram. Lab results for ESR, CRP, ANA, ENA, anti-DNA, anti-CCP antibodies, urate, and rheumatoid factor were also collected where available. Machine learning techniques were used to create a logistic regression model to classify referrals as inflammatory or non-inflammatory. Backward feature selection was used to enhance model performance by identifying features most predictive of referral state. Leave-one-out cross-validation was used to predict model performance. The models were subsequently evaluated on prospectively collected patient data from 20 new referrals seen after the creation of the triage tool, where data was coded blindly to patients’ diagnosis.

Results: In creation of the triage tool, 168 patient charts were used; 73 were classified as IA after being seen by a rheumatologist. The triage tool correctly classified 65 of 73 referrals as inflammatory using patient-reported information (model 1) (sensitivity 89%, specificity 52%). When the referral tool was reevaluated using laboratory markers in addition to the information obtained from the patient (model 2), 67 referrals were correctly classified as inflammatory (sensitivity 92%, specificity 63%). When model 1 was tested on 20 prospective patients, all 6 patients with IA were correctly classified (sensitivity 100%, specificity 52%). Model 2 correctly classified 14 patients (who had lab values present) including 5 patients with IA (sensitivity 83%, specificity 64%). The results are shown in table. Further prospective multi-site testing is ongoing.

Conclusion: A referral tool that can be entered into a database for computer assessment has good sensitivity to detect high priority referrals from data that can be obtained directly from patients and ascertained by non-healthcare workers. This may be of benefit in areas of limited resources and long waiting lists to see a rheumatologist.

ACR Meeting Abstracts - https://acrabstracts.org/abstract/computer-learning-artificial-intelligence-to-create-a-computer-based-triage-tool-classifying-referrals-as-inflammatory-or-non-inflammatory-requiring-only-patient-reported-information/