Abstract: TH-OR21
Prediction of Kidney Failure Using Multiple Data Domains in Glomerulonephropathy: A CureGN Study
Session Information
- Glomerular Diseases: Clinical and Translational Studies
November 02, 2023 | Location: Room 103, Pennsylvania Convention Center
Abstract Time: 04:30 PM - 04:39 PM
Category: Glomerular Diseases
- 1402 Glomerular Diseases: Clinical, Outcomes, and Trials
Authors
- Mariani, Laura H., University of Michigan, Ann Arbor, Michigan, United States
- Mansfield, Sarah, Arbor Research Collaborative for Health, Ann Arbor, Michigan, United States
- Smith, Abigail R., Arbor Research Collaborative for Health, Ann Arbor, Michigan, United States
- Tuttle, Katherine R., University of Washington, Seattle, Washington, United States
Group or Team Name
- CureGN Machine Learning Writing Group.
Background
Clinical risk factors do not fully predict kidney failure in patients with glomerular diseases. Using data from multiple domains in CureGN we applied machine learning to improve risk prediction and identify novel predictors of kidney failure.
Methods
Sequential ridge regression models using demographics (I), social determinants of health (SDOH; II), clinical (III), and pathology features (IV) were fitted to predict time to kidney failure (estimated glomerular filtration rate (eGFR) <15 mL/min/1.73m2 or treatment by dialysis or transplant). Discrimination was assessed by integrated area under the curve (iAUC); variables were ranked by absolute value of standardized coefficients.
Results
The kidney failure rate was 2.9 per 100 person years in 2,544 CureGN participants. 36 predictors were included (7 base, 2 socioeconomic, 27 clinical). Discrimination was similar between models I and II (iAUC=0.89) and higher for model III (iAUC=0.91). eGFR, urine protein to creatinine ratio, age, Black race, and FSGS were highly ranked across models I-III; the ranking of Black race was lower in model III (Figure). Medicaid as primary insurance, hypertension, renin angiotensin system inhibitor use, and serum albumin and urea nitrogen levels were highly ranked in model III. In model IV (n=670), interstitial fibrosis/tubular atrophy (IFTA), presence of tubular microcysts, and sclerotic glomeruli ranked as top predictors, displacing Black race, and reducing ranking of age and FSGS.
Conclusion
Machine learning improves prediction of kidney failure through incorporation of novel data domains. In CureGN, addition of these data improved prediction for kidney failure across conventional diagnostic categories and detected novel predictors that displaced traditional risk factors.
Varible rankings from models I-IV.
Funding
- NIDDK Support