Development and validation of a multivariable Prediction Model for Pre-diabetes and Diabetes using Easily Obtainable...

Area of Science:

• Endocrinology
• Public Health
• Medical Informatics

Background:

• Prevalence of pre-diabetes and diabetes is rising in the US, often alongside other chronic conditions.
• Hemoglobin A1c is a common diagnostic tool but can be inaccurate with co-existing diseases.
• There is a need for improved diagnostic methods for pre-diabetes and diabetes.

Purpose of the Study:

• To evaluate if readily available clinical data can enhance pre-diabetes and diabetes diagnosis compared to hemoglobin A1c alone.
• To develop and validate a predictive model using accessible clinical information.

Main Methods:

• Cross-sectional analysis of 13,800 participants from the US National Health and Nutrition Examination Survey (2005-2016).
• A machine learning model (gradient boosted decision tree) was used to estimate 2-hour glucose levels.
• Comparison of model performance (AUROC, predictive value, net benefit) against hemoglobin A1c and fasting plasma glucose.

Main Results:

• The 20-feature prediction model significantly outperformed hemoglobin A1c and fasting plasma glucose in diagnosing pre-diabetes and diabetes.
• Area-under-the-receiver-operating-curve (AUROC) improved from 0.66/0.71 to 0.77 for pre-diabetes and 0.87/0.88 to 0.91 for diabetes.
• Key features included standard vitals, fasting/non-fasting labs (glucose, insulin, lipids, liver function, kidney function), and the Poverty Ratio.

Conclusions:

• Incorporating easily accessible clinical data into electronic medical records can identify patients with undiagnosed pre-diabetes and diabetes.
• This approach may lead to earlier diagnosis and intervention, potentially preventing severe complications.
• The findings highlight the utility of a multivariable prediction model for improving diabetes screening.