Search

Background Patients with acute interstitial nephritis (AIN) can present without typical clinical features, leading to a delay in diagnosis and treatment. We therefore developed and validated a diagnostic model to identify patients at risk of AIN using variables from the electronic health record. Methods In patients who underwent a kidney biopsy at Yale University between 2013 and 2018, we tested the association of >150 variables with AIN, including demographics, comorbidities, vital signs and laboratory tests (training set 70%). We used least absolute shrinkage and selection operator methodology to select prebiopsy features associated with AIN. We performed area under the receiver operating characteristics curve (AUC) analysis with internal (held-out test set 30%) and external validation (Biopsy Biobank Cohort of Indiana). We tested the change in model performance after the addition of urine biomarkers in the Yale AIN study. Results We included 393 patients (AIN 22%) in the training set, 158 patients (AIN 27%) in the test set, 1118 patients (AIN 11%) in the validation set and 265 patients (AIN 11%) in the Yale AIN study. Variables in the selected model included serum creatinine {adjusted odds ratio [aOR] 2.31 [95% confidence interval (CI) 1.42–3.76]}, blood urea nitrogen:creatinine ratio [aOR 0.40 (95% CI 0.20–0.78)] and urine dipstick specific gravity [aOR 0.95 (95% CI 0.91–0.99)] and protein [aOR 0.39 (95% CI 0.23–0.68)]. This model showed an AUC of 0.73 (95% CI 0.64–0.81) in the test set, which was similar to the AUC in the external validation cohort [0.74 (95% CI 0.69–0.79)]. The AUC improved to 0.84 (95% CI 0.76–0.91) upon the addition of urine interleukin-9 and tumor necrosis factor-α. Conclusions We developed and validated a statistical model that showed a modest AUC for AIN diagnosis, which improved upon the addition of urine biomarkers. Future studies could evaluate this model and biomarkers to identify unrecognized cases of AIN. [ABSTRACT FROM AUTHOR]