Development and External Validation of Machine Learning Models for Diabetic Microvascular Complications:...

Area of Science:

• Biomedical data science
• Metabolomics
• Diabetology

Background:

• Diabetic kidney disease (DKD) and diabetic retinopathy (DR) are significant microvascular complications of diabetes.
• The kidney and eye share similar microvascular structures, making them susceptible to comparable metabolic changes during diabetes.

Purpose of the Study:

• To identify biomarkers for DKD and DR using machine learning (ML) and metabolic data in an Asian population.
• To improve DKD and DR detection models by incorporating ML beyond conventional risk factors.

Main Methods:

• Utilized ML algorithms (logistic regression with LASSO, gradient-boosting decision tree) on data from the Singapore Epidemiology of Eye Diseases study (n=2772).
• Analyzed 220 circulating metabolites and 19 risk factors to identify key variables for DKD and DR.
• Developed and externally validated DKD and DR detection models using the UK Biobank dataset (n=5843).

Main Results:

• Identified diabetes duration, insulin usage, age, and tyrosine as crucial factors for both DKD and DR.
• DKD was further associated with cardiovascular disease history, antihypertensive use, and specific metabolites (lactate, citrate, IDL cholesterol esters to total lipids ratio).
• DR was associated with HbA1c, blood glucose, pulse pressure, and alanine; ML models outperformed traditional logistic regression in both internal and external validation.

Conclusions:

• Diabetes duration, insulin usage, age, and tyrosine are vital for detecting DKD and DR.
• Integrating ML with big biomedical data facilitates biomarker discovery and enhances disease detection efficacy.