Deep Contrast Clustering Analysis to Distinguish Diabetic Complications in Elderly Chinese Patients

Abstract

Aim: To establish an innovative clustering method for predicting variable categories of diabetic complications in Chinese ≥ 65 with diabetes.

Materials and methods: We selected and extracted data from elderly patients with diabetes (n = 4980) from a medical examination group of 51,400 people followed up annually from 2014 to date in Kunshan, China. A deep contrast clustering approach was used to cluster and predict diabetic complications. The clustering approach was further validated using data from elderly patients with diabetes (n = 397) from one medical examination cohort of 20,000 people followed up yearly from 2014 to date in Beijing Jiuhua Hospital.

Results: The patients were clustered into 6 categories by analysing 20 indicators. Cluster 1-Heavy smoking and a high cardiovascular disease (CVD) risk; Cluster 2-High alcohol consumption, high aminotransferase levels, the highest risk of stroke complications, and a high fatty liver disease (FLD) risk; Cluster 3-High blood lipid levels and a risk of FLD and stroke complications; Cluster 4-Good health indicators and a low risk of FLD, stroke, and CVD complications; Cluster 5-Older age, higher uric acid concentration and creatinine level, and the highest risk of CVD complications; Cluster 6-Large waist circumference, high BMI, high blood pressure, and the highest risk of FLD complications. The gene for nonalcoholic fatty liver disease in cluster 2 had the highest risk coefficient. This was consistent with cluster 2, which had a higher FLD prevalence.

Conclusions: A new clustering method was developed from two large Chinese cohorts of older patients with diabetes, which may effectively predict complications by clustering into different categories.

Keywords: artificial intelligence; cluster analysis; diabetes classification; elderly diabetics.