P261 Early and accurate diagnosis of patients with axial spondyloarthritis using machine learning: a predictive analysis from electronic health records in the United Kingdom

Abstract

Background/Purpose: On an average, there is a delay of 6.7 years between symptom onset and diagnosis of axial spondyloarthritis (axSpA)1. Since traditional approaches to improving early axSpA identification have had limited success, predictive automated analyses on patient records may help alleviate the burden on healthcare providers2-5. Here, we report the results from a machine learning (ML) algorithm developed with UK electronic health records (EHRs) Clinical Practice Research Datalink (CPRD) data, to estimate the probability or likelihood of a pt being diagnosed with axSpA based on prior clinical indicators and pt history. Method(s): Primary care UK EHR data -CPRD GOLD was used to identify pts with axSpA and healthy controls (HC) (Figure). Pts aged >=18 years, with first diagnosis date of axSpA within the identification period (01-Jan-2005 to 31-Dec-2018) and fulfilling CPRD research acceptability criteria were included in the study. Data pertaining to clinical presentation, consultation, referral, test, and therapy history were extracted for individual pt prior to diagnosis of ax-SpA. A total of 5090 pts with axSpA satisfied the acceptability criteria. HC were randomly sampled to create a subset of one unique HC matched to each pt with axSpA, resulting in 5089 HC. ML usable features derived from the total population (pts with axSpA and HC) numbered 820. After using a further exclusion criterion for the pts with axSpA and HC who had >=1 out of 820 usable features, the final dataset included 7813 pts (3902 pts with axSpA and 3911 HC). This combined dataset was randomly split (67:33) into a train (N=5237) and a test (N=2576) dataset. A Random Forest (RF) model was trained on the train dataset. Cross-validation was performed for hyper-parameter tuning of the RF classifier. Once the model was trained, accuracy, precision, and F-1 scores were obtained with the test dataset. Result(s): The RF based algorithm resulted in a high level of accuracy (88.12%), with precision of 0.95 for pts with axSpA and 0.83 for HC (Figure). RF algorithm identified 89 best clinical predictors (out of 820 used as inputs) that differentiated between pt and HC such as: total number of tests, total number of referrals, first age of consultation, first symptom age, number of low back pain symptoms. The sensitivity of the model was 75.04% and positive predictive value was 80.88%. The specificity of the model was 0.96 and negative predictive value was 82.56%. Conclusion(s): The ML algorithm demonstrated a high level of accuracy and precision in the identification of possible cases of axSpA, which may be useful in reducing the delay in diagnosis. Previous studies have successfully demonstrated automated cohort identification of axSpA in large datasets, with only a few using ML based approaches for diagnosis from patient medical history. While the model supports previous work in axSpA2-5, it needs further validation in routine clinical practice and this exploration is ongoing.