Predictive ability of visit-to-visit glucose variability on diabetes complications

Abstract

Prevalence of Type 2 diabetes is increasing worldwide and it is expected that diabetes complications will follow. Glucose variability (GV) measures fluctuations of glucose homeostasis either by measuring intra-day, between-day, or visit-to-visit variations. There are limited studies that include the use of GV measures to predict diabetes complications using machine learning models. This study aimed to explore the predictive ability of different visit-to-visit GV measures (HbA1c-CV, HbA1c-SD, FPG-CV, FPG-SD, time-varying HbA1c and time-varying FPG) on three diabetes complications (CVD, DR, and CKD), by using Cox proportional hazard regression and machine learning models (random survival forest (RSF) and LTRC forest). This retrospective cohort study included 40,662 Type 2 diabetes patients from Ramathibodi Hospital, Thailand from 2010 to 2019. The results found that RSF models of DR complications had the highest C-index. All the GV measures also ranked the top in DR models. HbA1c-CV, HbA1c-SD, FPG-CV and FPG-SD were associated with all the three diabetes complications of interest (CVD, DR and CKD) whereas time-varying of HbA1c and FPG were only associated with CVD and DR. LTRC forest models showed the worst performance, with C-index ranging from 0.468-0.678 across all the models. In addition, it was observed that models developed using longitudinal dataset had higher C-index than baseline dataset. In conclusion, GV measures can be an additional marker for disease monitoring. HbA1c and FPG GV measures models had similar performance, which showed that FPG GV measures can be a good measure in resource-limited settings, when HbA1c cannot be performed. IMPLICATION OF THE THESIS: GV measures, in terms of HbA1c or FPG, can be used as additional monitoring parameters in diabetes management. Models can be deployed as clinical decision support tool/prompt by auto-calculation of GV and risk prediction of diabetes complications.