PENALIZED REGRESSION ANALYSIS IDENTIFIES CRITERIA AND NON-CRITERIA FEATURES THAT MAY INCREASE THE ACCURACY OF EXISTING SETS OF CRITERIA FOR CLASSIFYING SYSTEMIC LUPUS ERYTHEMATOSUS (SLE)

Background: The ACR-1997, SLICC-2012 and EULAR/ACR-2019 classification criteria have high sensitivity and specificity for SLE, yet they classify non-overlapping groups of patients suggesting that they can be supplemented with additional features to improve their diagnostic performance. Objectives: To identify criteria and non-criteria manifestations that are significantly associated with SLE in clinical practice and can be used to complement the existing sets of classification criteria. Methods: Individual items from all three classification criteria (ACR-1997, SLICC-2012, EULAR/ACR-2019) and non-criteria features were analyzed in a randomly selected sample of 800 adults diagnosed with SLE or control rheumatologic diseases (1:1 ratio). The classification performance of each set of criteria was analyzed in combination with complementary features; multivariable least absolute shrinkage and selection operator (LASSO) logistic regression was performed for feature selection. We calculated the diagnostic odds ratio (DOR) of the criteria and the additional features retained in each model. Results: Τhe EULAR/ACR-2019 and SLICC-2012 criteria have increased accuracy for SLE classification as compared to the ACR-1997 criteria (univariate DOR: 243.2 and 157.3 versus 78.8, respectively). In multivariable regression based on the ACR-1997 criteria, inclusion of additional features such as maculopapular rash, alopecia and hypocomplementemia significantly enhanced the model predictive capacity (area under the curve [AUC]: 0.95 versus 0.87 of the ACR-1997 criteria alone). Similar analysis based on the SLICC-2012 and EULAR/ACR-2019 criteria identified photosensitivity as an additional criterion significantly associated with SLE (multivariable DOR: 5.4 and 9.4, respectively). Accordingly, models including photosensitivity had superior predictive capacity over the criteria-only models (AUC: 0.94 versus 0.91 for SLICC-2012, 0.96 versus 0.91 for EULAR/ACR-2019). Furthermore, non-criteria features including Raynaud’s/livedo reticularis, anti-RNP antibodies, splenomegaly and myocarditis were independently associated with SLE thus enhancing further the predictive capacity of criteria-based models. Conclusion: We identified a number of criteria and non-criteria features which can be used in combination with the existing sets of criteria to increase classification of SLE patients in clinical practice. Photosensitivity could be considered as an additional feature to improve sensitivity of the recent classification criteria. Disclosure of Interests: Christina Adamichou: None declared, Irini Genitsaridi: None declared, Dionysis Nikolopoulos: None declared, Alessandra Bortoluzzi: None declared, Antonis Fanouriakis Paid instructor for: Paid instructor for Enorasis, Amgen, Speakers bureau: Paid speaker for Roche, Genesis Pharma, Mylan, Eleni Kalogiannaki: None declared, Emmanouil Papastefanakis: None declared, Irini Gergianaki: None declared, Prodromos Sidiropoulos: None declared, Dimitrios Boumpas: None declared, George Bertsias Grant/research support from: GSK, Consultant of: Novartis