OP03 Redesigning the patient management pathway through prediction of flares in axial spondyloarthritis using a machine learning approach to improve care

Abstract Background and aims Treatment of axial spondyloarthritis (axSpA) involves the management of flares which can be unpredictable. ASAS developed a data-driven definition of flares/disease worsening as an increase in ASDAS-CRP of at least 0.9 points, for use in axSpA clinical trials (1). With the current pressures on clinic capacity, there is a challenge to see flare patients in a timely manner. Artificial intelligence (AI) and natural language processing (NLP) methods provide algorithms for learning systems to recognise disease-related terms and classify clinical phenotypes using large datasets, which may support early identification of patients with a risk of having a flare. In previous work, we presented an approach using NLP to identify patients who may have rheumatic disease with predictors to discriminate patients with diagnosed inflammatory arthritis from controls with non-inflammatory arthritis. With the same aim, we extended the panel of predictors by combining clinical and laboratory data to predict flares in axSpA. Methods 282 axSpA patients seen between January 2016 to May 2024 were included in the study. There are 100 flare and 182 non-flare patients studied. We utilised patient blood test results, demographic information, electronic patient-reported outcomes scores, comorbidity, weight and height. We developed a time series dataset and applied Long Short-Term Memory (LSTM) networks and neural networks (NN) to forecast the risk of flare before their future clinics. Results We conducted predictive modelling to forecast flare and non-flare events using data collected prior to patients upcoming clinic appointments. The dataset was split into training and testing sets based on observations taken 2 months before the clinic visit. Our LSTM model achieved 90.25% accuracy, 0.89 sensitivity, 0.96 specificity, 0.93 precision and 0.98 Area Under the Receiver Operating Characteristic Curve (AUROC) predicting flare and non-flare events 2 months before the clinics. The results are summarized in Table 1. OP03 Table 1.Performance of Long Short-Term Memory (LSTM) and neural networks (NN) on data from 2 months before the clinic visit.AVG ACCURACYAVG SENSITIVITYAVG SPECIFICITYAVG PRECISIONAVG AUROCLSTM (2 months before the next clinics)90.25%0.890.960.930.98NEURAL NETWORK (2 months before the next clinics)71.90%0.910.010.610.5 Conclusions The presented combination of clinical and laboratory data shows promising results for the prediction of flare in axSpA and may in the future assist clinicians in seeing the right patient in the right clinic. A study is planned not only to confirm the results but also to refine the machine learning (ML) models developed and specify disease attributes. In the service redesign, patients identified as at risk of flares will be booked and seen in routine clinics at the right time to prevent flares. This process in the long-term will reduce the reliance on flare and urgent clinics. References 1. Molto A, et al. Ann Rheum Dis. 2018 Jan;77(1):124–7.