Background/Purpose: Lupus nephritis (LN) outcomes are affected not only by patient biology but also by patient, provider, and system factors. One way to improve outcomes in LN is to use a systems approach to stratify all patients in a population by their risk for LN treatment failure. We have demonstrated that traditional and novel biomarkers obtained at the time of renal biopsy can be used to create random forest machine learning models of one-year treatment outcomes. However, to predict system-level outcomes for care coordination, one must leverage available electronic health record (EHR) data to model outcomes. This project was designed to determine the effectiveness of using currently available EHR data to create random forest models predictive of one-year outcomes in LN.

Methods: 212 LN patients from one academic prospective cohort and two clinical trials of lupus nephritis were selected as follows. LN was defined by International Society for Nephrology/Renal Pathology Society (ISN/RPS) renal biopsy classification within two months of baseline data collection. All patients were started on mycophenolate mofetil as induction therapy. The following data were collected at the start of induction: age, sex, race, biopsy class (ISN/RPS), C3, C4, and DNA antibody levels, UPrCr, and estimated glomerular filtration rate (eGFR, Chronic Kidney Disease Epidemiology Collaboration formula). One-year complete response outcome was calculated from baseline and one year eGFR and UPrCR using the ACR renal response criteria modified for the LN and Rituximab trial. Random forest models were created in using the baseline EHR variables above to predict a one-year outcome of complete response versus partial response or treatment failure. A separate test set from 1/3 of the data was created for reporting the receiving operator characteristics area under the curve (ROC AUC).

Results: The model predicted the outcome in the test set with an ROC AUC of 0.74.

Conclusion: These results demonstrate that treatment response in patients on mycophenolate mofetil induction therapy can be predicted with a reasonable level of accuracy using data available in the EHR. This level of accuracy is sufficient to prioritize care coordination efforts to patients most at risk. The ideal implementation of this model would calculate the risk score in near real time in the EHR. Because the ISN/RPS classification is not typically reported in discreet fields, natural language processing would be necessary to extract these data. To have impact at a population level, the risk score must be visible in population level reports or dashboards so that high risk patients can be targeted for assistance to reduce patient, provider, and system barriers to good outcomes.

« Back to 2016 ACR/ARHP Annual Meeting

ACR Meeting Abstracts - https://acrabstracts.org/abstract/random-forest-models-using-electronic-health-record-data-are-predictive-of-one-year-outcomes-in-lupus-nephritis-patients-taking-mycophenolate-mofetil-induction-therapy/