LBA:01:18 Machine learning-driven systemic lupus erythematosus activity evaluation system based on conventional laboratory indicators

Objectives The disease activity assessment of systemic lupus erythematosus (SLE) mainly relies on the SLEDAI score, but its evaluation indicators (such as dsDNA antibodies) are difficult to detect accurately, and there are limitations in the application of SLEDAI score to evaluate SLE activity at the grassroots level. Conventional laboratory indicators have potential value in SLE activity evaluation due to their simple detection methods and unified evaluation criteria. Therefore, this study proposes a data-driven evaluation framework to develop a routine laboratory index evaluation system for SLE disease assessment activities through systematic analysis of routine laboratory indicators.Methods In this study, Bayesian linear regression regularization optimization (BLR-ROI) method was used to process missing data, and the weight coefficients (Beta) corresponding to each feature were obtained by combining recursive feature elimination (RFE) and linear kernel function, and 20 key features with a contribution rate >1% were selected from 58 routine laboratory indicators of 709 patients, and the non-SLEDAI routine laboratory feature set XBest was formed, and 1-9 points were assigned according to the linear normalization of the contribution rate. The characteristic clustering centers of the active group and the inactive group were calculated, and the midpoint of the mean of the two groups was taken as the judgment threshold. Finally, the performance accuracy of the model was evaluated by 5-fold cross-validation, and the sensitivity (TP/(TP+FN)) and specificity (TN/(TN+FP)) were calculated.Results The 20 features screened by RFE were constructed based on the contribution rate, including albumin (9.29%), neutrophil percentage (8.59%), and free cholesterol (8.23%). The classification accuracy of the support vector machine model reached 76.38%(sensitivity 0.615, specificity 0.791), which was significantly better than that of the full-feature model (71.46%).The conventional laboratory index scoring system achieved the best discriminant effect (p