Comparing Machine Learning and Penalized Regression for Predicting Diabetic Kidney Disease Progression: Evidence from the Chronic Renal Insufficiency Cohort (CRIC) Study

• Activity Number: 205 - Applications of Machine Learning
• Type: Contributed
• Date/Time: Tuesday, August 4, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Jing Zhang* and Tobias Fuhrer and Brian Kwan and Daniel Montemayor and Kumar Sharma and Loki Natarajan
• Companies: Moores Cancer Center, University of California, San Diego and Institute of Molecular Systems Biology, ETH Zurich, Zurich, Switzerland and University of California, San Diego and Department of Medicine, University of Texas Health Science Center at San Antonio and Department of Medicine, University of Texas Health Science Center at San Antonio and University of California, San Diego
• Keywords: diabetic kidney disease; penalized regression; lasso; random forest

Abstract

There is an urgent need to identify novel biomarkers that can predict and elucidate mechanisms of diabetic kidney disease (DKD). Urine samples from 995 CRIC patients were assayed for relative metabolite abundance yielding 1899 annotated features. Stringent filtering criteria were applied to eliminate noisy features, resulting in 698 reliably measured features. We fit prognostic models for kidney function decline (defined as eGFR slope), using penalized (Lasso) and Random forest (RF) models, with metabolites and clinical predictors (demographics, BMI, blood pressure, HbA1c, eGFR, albuminuria). The models with lowest prediction error were further evaluated on the time-to-ESRD outcome via c-statistics. Five-fold cross validation was used to obtain the c-statistics (median, 95%CI). The eGFR slope models selected 9 - 122 features depending on lasso penalty and RF variable importance. The best ESRD model with 20 metabolites & 9 clinical factors, had a c-statistic of 0.85 (0.85, 0.86). Pathway analysis revealed that prognostic metabolites were involved in cell signaling, energy storage etc. Modern statistical methods applied to untargeted metabolomics can reveal novel insights in DKD.