Artificial intelligence-based diagnosis of diabetic kidney disease using urinary VOC biosensor data

Area of Science:

• Nephrology
• Biomarker Discovery
• Artificial Intelligence in Medicine

Background:

• Diabetic kidney disease (DKD) is a major cause of chronic kidney disease globally.
• Current diagnostic methods for DKD are invasive (renal biopsy) or rely on indirect biomarkers.
• There is a need for non-invasive diagnostic tools for DKD.

Purpose of the Study:

• To evaluate the feasibility of using urinary volatile organic compound (VOC) profiling for non-invasive DKD classification.
• To assess the performance of machine learning models in identifying DKD from VOC profiles.
• To explore the potential of VOC analysis as a diagnostic triage tool.

Main Methods:

• Collected urine samples from 127 participants across four groups: healthy controls, diabetes without nephropathy, biopsy-confirmed DKD, and primary nephrotic syndromes.
• Analyzed samples using a chemiresistive VOC biosensor, extracting over 15,000 signal-derived features.
• Trained and validated four machine learning classifiers (Random Forest, SVM, k-NN, Naïve Bayes) using SMOTE for class balancing and stratified data.

Main Results:

• The Random Forest model achieved the highest performance: 86% accuracy, 0.91 precision, 0.86 recall, 0.86 F1-score, and 0.95 AUC.
• K-fold cross-validation confirmed the model's robustness and generalizability.
• Random Forest effectively differentiated DKD from other diabetic and non-diabetic glomerular diseases.

Conclusions:

• Urinary VOC profiling combined with machine learning offers a proof-of-concept for non-invasive DKD diagnosis.
• The Random Forest model shows promise as a triage tool to identify DKD.
• This approach may reduce the need for biopsies and facilitate earlier DKD detection in clinical practice.