Predicting rapid decline in kidney function among type 2 diabetes patients: A machine learning approach.

Area of Science:

• Nephrology
• Endocrinology
• Data Science

Background:

• Diabetic kidney disease (DKD) is a common complication of type 2 diabetes (T2D).
• A subset of DKD patients experience rapid decline (RD) in kidney function, increasing dialysis risk.
• Current biomarkers like albuminuria are insufficient for predicting RD, necessitating new predictive models.

Purpose of the Study:

• To develop a machine learning model for predicting RD in T2D patients.
• To identify key laboratory tests that contribute to the prediction of RD.
• To understand the mechanisms underlying RD in DKD through comprehensive laboratory test analysis.

Main Methods:

• A machine learning model was developed to predict RD within one year, using estimated glomerular filtration rate (eGFR) as an indicator.
• Recursive feature elimination with cross-validation (RFECV) was employed to select the most predictive laboratory tests.
• The model was trained and validated on 1202 laboratory tests from 3438 T2D patients at the University of Tokyo Hospital.

Main Results:

• The 8-feature model achieved a receiver operating characteristic area under the curve (ROC-AUC) of 0.820.
• RFECV identified 7 key laboratory tests (MCH, γ-GTP, Cre, HbA1c, HDL-C, eGFR, and Hct) contributing to RD prediction.
• The model's performance (ROC-AUC 0.820) surpasses previous studies (ROC-AUC 0.775).

Conclusions:

• The developed model accurately predicts rapid kidney function decline in T2D patients.
• The identified laboratory tests may serve as novel biomarkers for DKD progression.
• Physicians can utilize this model to focus on interventions that inhibit kidney damage progression.