Development of a 5-year risk prediction model for type 2 diabetes in individuals with incident HbA1c-defined pre-diabetes in Denmark

Abstract

Introduction Pre-diabetes increases the risk of type 2 diabetes, but data are sparse on predictors in a population-based clinical setting. We aimed to develop and validate prediction models for 5-year risks of progressing to type 2 diabetes among individuals with incident HbA1c-defined pre-diabetes. Research design and methods In this population-based cohort study, we used data from the Danish National Health Survey (DNHS; n=486 495), linked to healthcare registries and nationwide laboratory data in 2012-2018. We included individuals with a first HbA1c value of 42-47 mmol/mol (6.0%-6.4%), without prior indications of diabetes. To estimate individual 5-year cumulative incidences of type 2 diabetes (HbA1c ≥48 mmol/mol (6.5%)), Fine-Gray survival models were fitted in random 80% development samples and validated in 20% validation samples. Potential predictors were HbA1c, demographics, prescriptions, comorbidities, socioeconomic factors, and self-rated lifestyle. Results Among 335 297 (68.9%) participants in DNHS with HbA1c measurements, 26 007 had pre-diabetes and were included in the study. Median HbA1c was 43.0 mmol/mol (IQR 42.0-44.0 mmol/mol, 6.1% (IQR 6.0%-6.2%)), median age was 69.6 years (IQR 61.0-77.1 years), and 51.9% were women. During a median follow-up of 2.7 years, 11.8% progressed to type 2 diabetes and 10.1% died. The final prediction model included HbA1c, age, sex, body mass index (BMI), any antihypertensive drug use, pancreatic disease, cancer, self-reported diet, doctor's advice to lose weight or change dietary habits, having someone to talk to, and self-rated health. In the validation sample, the 5-year area under the curve was 72.7 (95% CI 71.2 to 74.3), and the model was well calibrated. Conclusions In addition to well-known pre-diabetes predictors such as age, sex, and BMI, we found that measures of self-rated lifestyle, health, and social support are important and modifiable predictors for diabetes. Our model had an acceptable discriminative ability and was well calibrated.