A non-invasive urinary diagnostic signature for diabetic kidney disease revealed by machine learning and single-cell analysis
Fig 5
Machine learning-based discovery and validation of diagnostic biomarkers for DKD.
(A) LASSO regression coefficient profile of genes differentially expressed in early vs. advanced DKD urinary cells. (B-G) Receiver operating characteristic (ROC) curves evaluating the diagnostic performance of PDK4, RHCG, and FBP1 in distinguishing DKD from controls in the training set (n = 40 DKD, 21 controls) and independent validation sets (n = 30 DKD, 42 controls). (H-I) Performance evaluation of multi-gene diagnostic models built using four machine learning algorithms in the training set (H) and independent validation set (n = 27 DKD, 9 controls) (I). (J) Box plots showing the expression levels of PDK4, RHCG, and FBP1 in control versus DKD groups (n = 21 controls, 40 DKD). Statistical significance was determined using the Wilcoxon rank-sum test; p-values are unadjusted. (*P < 0.05, **P < 0.01, ***P < 0.001). (K) Scatter plot with regression line showing the correlation between FBP1 expression and estimated glomerular filtration rate (eGFR) in DKD patients from the Nephroseq database.