In Rheumatology (Oxford, England)
BACKGROUND : The ability to identify lupus patients in High Disease Activity Status (HDAS) without knowledge of the SLEDAI could have application in selection of patients for treatment escalation or enrolment in trials. We sought to generate an algorithm that could calculate via model fitting the presence of HDAS using simple demographic and laboratory values.
METHODS : We examined the association of High Disease Activity (HDA) with demographic and laboratory parameters using prospectively collected data. An HDA visit is recorded when SLEDAI-2K ≥10. We utilised the use of combinatorial search to find algorithms to build a mathematical model predictive of HDA. Performance of each algorithm was evaluated using multi-class area under receiver operating characteristics (mAUROC) and the final model was compared with the Naïve Bayes Classifier, and analysed using the confusion matrix for accuracy and misclassification rate.
RESULTS : Data on 286 patients, followed for a median of 5.1 years were studied for a total of 5,680 visits. Sixteen laboratory parameters were found to be significantly associated with HDA. A total of 216 algorithms were evaluated and final algorithm chosen was based on 7 pathology measures and 3 demographic variables. It has an accuracy of 88.6% and misclassification rate of 11.4%. When compared with the Naïve Bayes Classifier (AUC = 0.663), our algorithm has a better accuracy with AUC = 0.829.
CONCLUSION : This study shows that building an accurate model to calculate HDA using routinely available clinical parameters is feasible. Future studies to independently validate the algorithm will be needed to confirm its predictive performance.
Hoi Alberta, Nim Hieu T, Koelmeyer Rachel, Sun Ying, Kao Amy, Gunther Oliver, Morand Eric
High Disease Activity Status, Systemic lupus erythematosus, machine learning