Development of a lupus nephritis suboptimal response prediction tool using renal histopathological and clinical laboratory variables at the time of diagnosis.

objectiveLupus nephritis (LN) is an immune complex-mediated glomerular and tubulointerstitial disease in patients with SLE. Prediction of outcomes at the onset of LN diagnosis can guide decisions regarding intensity of monitoring and therapy for treatment success. Currently, no machine learning model of outcomes exists. Several outcomes modelling works have used univariate or linear modelling but were limited by the disease heterogeneity. We hypothesised that a combination of renal pathology results and routine clinical laboratory data could be used to develop and to cross-validate a clinically meaningful machine learning early decision support tool that predicts LN outcomes at approximately 1 year.

methodsTo address this hypothesis, patients with LN from a prospective longitudinal registry at the Medical University of South Carolina enrolled between 2003 and 2017 were identified if they had renal biopsies with International Society of Nephrology/Renal Pathology Society pathological classification. Clinical laboratory values at the time of diagnosis and outcome variables at approximately 1 year were recorded. Machine learning models were developed and cross-validated to predict suboptimal response.

resultsFive machine learning models predicted suboptimal response status in 10 times cross-validation with receiver operating characteristics area under the curve values >0.78. The most predictive variables were interstitial inflammation, interstitial fibrosis, activity score and chronicity score from renal pathology and urine protein-to-creatinine ratio, white blood cell count and haemoglobin from the clinical laboratories. A web-based tool was created for clinicians to enter these baseline clinical laboratory and histopathology variables to produce a probability score of suboptimal response.

conclusionGiven the heterogeneity of disease presentation in LN, it is important that risk prediction models incorporate several data elements. This report provides for the first time a clinical proof-of-concept tool that uses the five most predictive models and simplifies understanding of them through a web-based application.