Kidney Week

Abstract: FR-OR02

A Parsimonious Model for Diagnosis of Biopsy-Proven Acute Interstitial Nephritis Using Electronic Health Record Data

Session Information

• AKI Predictors and Outcomes: Research Abstracts
  October 23, 2020 | Location: Simulive
  Abstract Time: 05:00 PM - 07:00 PM

Category: Acute Kidney Injury

• 102 AKI: Clinical, Outcomes, and Trials

Authors

• Moledina, Dennis G., Yale University School of Medicine, New Haven, Connecticut, United States

Dennis G. Moledina, MD, PhD



Dr. Moledina is an Assistant Professor of Medicine (Nephrology) at Yale School of Medicine, New Haven, CT. He is board-certified in Nephrology and Internal Medicine. He received his MD from Saurashtra University in India and PhD in Investigative Medicine from Yale University. Dr. Moledina is an NIH-funded clinical and translational researcher focused on improving diagnosis and management of acute kidney injury with specific focus on acute interstitial nephritis.

• Yamamoto, Yu, Yale University School of Medicine, New Haven, Connecticut, United States

Yu Yamamoto,

• Parikh, Chirag R., Johns Hopkins Medicine, Baltimore, Maryland, United States

Chirag R. Parikh,

• Wilson, Francis Perry, Yale University School of Medicine, New Haven, Connecticut, United States

Francis Perry Wilson,

Background

Due to its atypical clinical features and difficulty in establishing diagnosis without a biopsy, acute interstitial nephritis (AIN) diagnosis is delayed or missed. We developed a predictive model for AIN using clinical data from all patients who underwent a kidney biopsy available through the electronic health record.

Methods

We obtained data on all patients who underwent a native kidney biopsy at two centers between 2013-18 and obtained corresponding information of demographics, comorbidities, and all laboratory tests collected up to one year before biopsy. We used least absolute shrinkage and selection operator (LASSO) method to select features associated with AIN and performed area under received operating characteristics curve (AUC) analysis in temporally-split training (70%) and test (30%) sets. We also applied this model to a separate cohort of kidney biopsies with AIN diagnosis adjudicated by 3 pathologists and compared it to the clinician’s prebiopsy impression of AIN obtained through chart review.

Results

Among 551 patients who underwent native kidney biopsies, 60 (11%) had AIN on clinical pathology diagnosis. We evaluated 163 potential features for their association with AIN. The five features with the highest AUC were last creatinine at the time of biopsy (AUC, 0.73), BUN to creatinine ratio (0.70), urine specific gravity before biopsy (0.67), serum bicarbonate (0.62), and urine protein (0.62). The top 4 variables picked using LASSO had an AUC of 0.76 in the test set (table). Applying this model to a separate cohort of participants with adjudicated AIN, we noted an AUC of 0.80 (0.73, 0.87), which was higher than the clinician’s pre-biopsy impression of AIN (0.61 (0.52, 0.70), P<0.001).

Conclusion

We noted four variables associated with AIN and the model containing these showed a modest AUC but was an improvement on clinician’s pre-biopsy impression of AIN.

Features associated with AIN

AUC, area under receiver operating characteristic curve; OR, odds ratio; CI, confidence interval AUC training set=0.79

Funding

• NIDDK Support