Machine learning models rank global early-onset diabetes exposures
High body mass index was the leading predictive exposure in three machine learning models and remained positively associated with early-onset type 2 diabetes incidence after adjustment for year and country characteristics in an ecological analysis of 204 countries and territories from 1990 to 2021.
The ecological study, published in Diabetes, Obesity and Metabolism, identified high body mass index (BMI) and low whole-grain intake as the two leading exposures across three machine learning models. Of these , only high BMI remained significantly associated with incidence after adjustment for both year and country fixed effects.
“These findings should be interpreted as population-level relationships between country-level exposure patterns and early-onset type 2 diabetes incidence rather than as evidence of individual-level risk or causality,” the authors wrote.
Model development and evaluation
Researchers from Harbin Medical University in Heilongjiang, China, used country-level aggregate data from the Global Burden of Disease Study 2021. They defined early-onset type 2 diabetes (EOT2D) as type 2 diabetes occurring at ages 15 to 39 years. Exposure was measured using summary exposure values, risk-weighted population measures ranging from 0 to 100, with higher values indicating greater adverse exposure.
Researchers screened 58 candidate exposure indicators for epidemiological relevance, potential applicability to population-level prevention, and collinearity, retaining 18 for machine learning analyses.
The researchers trained Elastic Net, Random Forest, and extreme gradient boosting (XGBoost) models using 1990 to 2012 data, selected hyperparameters using 2013 to 2016 data, and tested performance using 2017 to 2021 data. Standardization and collinearity screening used only training-period data to prevent information from later periods from entering model development.
XGBoost performed best in the validation and test periods. Test-set coefficient of determination (R²) values were 0.838 for XGBoost, 0.769 for Random Forest, and 0.559 for Elastic Net. XGBoost also had the lowest test-set prediction errors, with a root mean square error of 67.67 and mean absolute error of 47.697.
All three models showed lower predictive performance in the validation and test periods than during training. The authors noted that the declines for Random Forest and XGBoost suggested some overfitting.
Exposure rankings and adjusted associations
Across 300 country-level bootstrap resamples, XGBoost maintained the best temporal predictive performance, with a median test-set R² of 0.827. Robust rank aggregation, which combined exposure rankings across algorithms, retained only high BMI and low whole-grain intake after correction for multiple comparisons.
Supplementary SHapley Additive exPlanations (SHAP) analyses ranked high BMI first and low whole-grain intake second in all three models during both validation and testing. Other exposures showed less consistent rankings.
In the primary regression model, which included both selected exposures and adjusted for common year-specific influences, both exposures were positively associated with incidence. With exposure measures standardized and the natural logarithm of incidence as the outcome, β coefficients were 0.28 for high BMI and 0.158 for low whole-grain intake.
After additional adjustment for stable country characteristics, the association with high BMI weakened but remained statistically significant (β = 0.109). The coefficient for low whole-grain intake fell to 0.014 and was no longer statistically significant.
The authors noted that machine learning models could capture differences between countries, whereas adjustment for country fixed effects emphasized changes within countries over time.
Geographic patterns and limitations
EOT2D incidence increased globally, with the highest rates remaining concentrated in Pacific Island countries. In 2021, incidence was highest in the Cook Islands at 896.08 per 100,000 population, followed by Niue at 869.03 and Palau at 846.76. From 1990 to 2021, Egypt, Greenland, and Afghanistan had the largest estimated annual percentage increases, at approximately 5%, 5%, and 4%, respectively.
An exploratory descriptive index combining average exposure to high BMI and low whole-grain intake from 1990 to 2021 showed higher joint exposure in selected parts of North America, Oceania, Europe, and North Africa. Low levels were more common across sub-Saharan Africa and large parts of Asia.
Limitations included the use of country-level aggregate data, which precluded inference about patient-level risk or causal effects. Incidence and exposure estimates depended on underlying data availability and modeling assumptions. Neither regression approach could eliminate confounding by unmeasured factors that changed over time, including socioeconomic conditions, food environments, health care access, and diagnostic capacity.
“Future studies using individual-level and subnational longitudinal data are needed to further clarify temporal relationships, potential causal pathways and context-specific opportunities for preventing EOT2D in younger populations,” the authors wrote.
The authors declared no conflicts of interest.
AACE Endocrine AI is published by Conexiant under a license arrangement with the American Association of Clinical Endocrinology, Inc. (AACE®). The ideas and opinions expressed in AACE Endocrine AI do not necessarily reflect those of Conexiant or AACE. For more information, see Policies.