Kidney Week

Abstract: FR-PO1112

Insights from Machine Learning Analysis of Collated Lupus Nephritis Trials

Session Information

• Glomerular Diseases: Clinical, Outcomes, Trials - II
  October 26, 2018 | Location: Exhibit Hall, San Diego Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Glomerular Diseases

• 1203 Glomerular Diseases: Clinical, Outcomes, and Trials

Authors

• Gomez Mendez, Liliana Michelle, University of California San Francisco, San Francisco, California, United States Minor Outlying Islands

Liliana Michelle Gomez Mendez,

• Dai, Jian, Genentech, South San Francisco, California, United States

Jian Dai,

• Kutarnia, Jason, Mitre, Bedford, Massachusetts, United States

Jason Kutarnia,

• Howser, Scott, DataRobot, Boston, Massachusetts, United States

Scott Howser,

• Schuren, Jay, DataRobot, Boston, Massachusetts, United States

Jay Schuren,

• Mcclintock, Dana, Genentech, Inc, South San Francisco, California, United States

Dana Mcclintock,

• Cascino, Matthew, Genentech, Inc., South San Francisco, California, United States

Matthew Cascino,

• Prunotto, Marco, F. Hoffmann-La Roche Ltd., Basel, Switzerland

Marco Prunotto,

Background

Machine learning (ML) analyses are being increasingly used in diverse fields of medical research to identify parameters associated with disease progression and/or outcome. We tested this method for identifying variables associated with complete renal response (CRR) in lupus nephritis (LN).

Methods

Data collated from two clinical trials, LUNAR (NCT00282347)¹ and BELONG (NCT00626197)², compared the addition of rituximab¹ or ocrelizumab² vs. placebo in addition to background therapy for the treatment of LN. We deployed supervised ML method of symbolic regression using Eureqa® software to analyze variables for CRR in these studies. For patients with baseline creatinine >1, CRR was defined as urine protein/creatinine ratio (UPCR) <0.5 and creatinine <1 at 1 year. CRR for patients with baseline creatinine <1 was UPCR <0.5 and creatinine <125% of baseline at 1 year. Logistic regression was used to validate the association between identified variables and CRR found via ML analyses.

Results

We found that unbiased ML analysis identified variables known to be associated with CRR, such as baseline UPCR and eGFR. Furthermore, ML analysis identified other variables to be associated with CRR, including systolic blood pressure (BP) at baseline and Week 16, and positive anti-RNP antibody at baseline. Logistic regression determined that every 10 mmHg increment in systolic BP had an odds ratio (OR) of 0.84 (95% CI: 0.73, 0.96) for CRR at 1 year. At Week 16, every 10 mmHg increment in systolic BP had an OR of 0.82 (95% CI: 0.7, 0.95) for CRR at 1 year. Logistic regression showed an OR of 2.1 (95% CI: 1.3, 3.5) for CRR at 1 year with anti-RNP antibody positivity at baseline (n=182), adjusting for treatment received, baseline proteinuria, and eGFR.

Conclusion

In conclusion, ML algorithms can be successfully used in identifying variables associated with clinical response, especially in a complex and heterogeneous disease such as LN. These analyses highlighted systolic BP at baseline and Week 16, as well as baseline anti-RNP antibody positivity, to be associated with CRR. The association between anti-RNP antibody at baseline and CRR at 1 year warrants further investigation.

1. Rovin et al., Arthritis Rheum 2012; 64, 1215-26.
2. Mysler et al., Arthritis Rheum 2013; 65, 2368-79.

Funding

• Commercial Support – Genentech, Inc., South San Francisco, CA, USA