Kidney Week

Abstract: TH-PO612

Machine Learning-Derived Predictors of Clinical Outcomes in CureGN Participants with FSGS

Session Information

• Membranous Nephropathy, FSGS, and Minimal Change Disease
  October 24, 2024 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Glomerular Diseases

• 1402 Glomerular Diseases: Clinical, Outcomes, and Therapeutics

Authors

• Elliott, Mark, The University of British Columbia Faculty of Medicine, Vancouver, British Columbia, Canada

Mark Elliott, MD

• Helmuth, Margaret, University of Michigan, Ann Arbor, Michigan, United States

Margaret Helmuth, MA

• Smith, Abigail R., Northwestern University, Evanston, Illinois, United States

Abigail R. Smith, PhD

• D'Agati, Vivette D., Columbia University Irving Medical Center, New York, New York, United States

Vivette D. D'Agati, MD

• Barisoni, Laura, Duke University, Durham, North Carolina, United States

Laura Barisoni, MD

• Sanna-Cherchi, Simone, Columbia University Irving Medical Center, New York, New York, United States

Simone Sanna-Cherchi, MD

• Cattran, Daniel C., Toronto General Hospital, Toronto, Ontario, Canada

Daniel C. Cattran, MD

Group or Team Name

• CureGN.

Background

Focal segmental glomerulosclerosis (FSGS) is a common disease pattern with variable clinical presentations, outcomes, and etiologies, including immune-mediated, adaptive, and genetic. Much of the clinical heterogeneity remains unexplained.

Methods

722 participants with FSGS have been enrolled with 512 having whole genome sequencing (WGS) data, 276 with centralized biopsy scoring for conventional morphologic parameters, and 226 having both WGS and pathology scoring. Sequential ridge regression models were performed using: 1) baseline demographic, 2) social determinants of health, clinical, and genetic, 3) pathology morphologic features to predict time to kidney failure defined as two eGFR measurements <15mL/min, dialysis, or kidney transplantation. Ridge regression allows complex modelling while avoiding overfitting. Missing variables were imputed where possible. Discrimination was assessed using integrated area under the curve (iAUC) and variables were ranked by absolute value of standardized coefficients.

Results

Among participants with WGS data, 92 had a high-risk APOL1 genotype and 32 had monogenic kidney disorders. Model discrimination improved with the stepwise addition of features from model 1 to 3 (iAUC = 0.86, 0.93, 0.95 respectively). Among the 31 most predictive features 14 were protective (1 demographic, 6 clinical, 7 morphologic) and 17 were risk factors (2 demographic, 7 clinical, 1 genetic, 7 morphologic). The top ranked protective factors included high eGFR, high serum albumin, low rates of interstitial fibrosis and tubular atrophy (IFTA), high calcium, diffuse mesangial hypercellularity, high hemoglobin and sodium, tip variant FSGS, and no inflammation in areas of IFTA. Top ranked risk factors included high levels of IFTA, high proteinuria, tubular microcystic changes, hypertension, collapsing variant FSGS, inflammation in areas of IFTA, high levels of segmental and globally sclerosed glomeruli, thrombotic diagnosis, high potassium, edema, high-risk APOL1 genotype, private health insurance, high levels of arteriosclerosis, black race, use of RAAS inhibitors, and young age at onset.

Conclusion

Machine learning methodologies that integrate a broad range of clinical, genetic, and pathology data allow for the identification of parameters that are predictive of clinical outcome in FSGS.

Funding

• NIDDK Support