Machine learning with validation to detect diabetic microvascular complications using clinical and metabolomics data

ABSTRACT AIMS Using machine learning integrated with clinical and metabolomic data to identify biomarkers associated with diabetic kidney disease (DKD) and diabetic retinopathy (DR), and to improve the performance of DKD/DR detection models beyond traditional risk factors. METHODS We examined a population-based cross-sectional sample of 2,772 adults with type 1 or type 2 diabetes from Singapore Epidemiology of Eye Diseases study (SEED, 2004-2011). LASSO logistic regression (LASSO) and gradient boosting decision tree (GBDT) were used to select markers of prevalent DKD (defined as an eGFR < 60ml/min/1.73m2) and prevalent DR (defined as an ETDRS severity level ≥ 20) from an expanded set of 19 established risk factors and 220 NMR-quantified circulating metabolites. Risk assessment models were developed based on the variable selection results and externally validated in UK Biobank (n=5,843, 2007-2010). Model performance (AUC with 95% CI, sensitivity, and specificity) of machine learning was compared to that of traditional logistic regression adjusted for age, gender, diabetes duration, HbA 1c %, systolic BP, and BMI. RESULTS SEED participants had a median age of 61.7 years, with 49.1% female, 20.2% having DKD, and 25.4% having DR. UK Biobank participants had a median age of 61.0 years, with 39.2% female, 6.4% having DKD, and 5.7% having DR. Both algorithms identified diabetes duration, insulin usage, age, and tyrosine as the most important factors of both DKD and DR. DKD was additionally associated with CVD, hypertension medication, and three metabolites (lactate, citrate, and cholesterol esters to total lipids ratio in intermediate-density-lipoprotein); While DR was additionally associated with HbA 1c , blood glucose, pulse pressure, and alanine. Machine-learned models for DKD and DR detection outperformed traditional logistic regression in both internal (AUC: 0.832-0.838 vs. 0.743 for DKD, and 0.779-0.790 vs. 0.764 for DR) and external validation (AUC: 0.737-0.790 vs. 0.692 for DKD, and 0.778 vs. 0.760 for DR). CONCLUSIONS Machine-learned biomarkers suggested insulin resistance to be a primary factor associated with diabetic microvascular complications. Integrating machine learning with biomedical big data enabled biomarker discovery from a wide range of correlated variables, which may facilitate our understanding of the disease mechanisms and improve disease screening.