Limited utility of diet quality scores in anthropometric prediabetes and diabetes screening models

Background

Undiagnosed prediabetes and type 2 diabetes are major public health challenges. Machine learning tools utilizing anthropometrics effectively predict dysglycemia, but poor diet is a fundamental driver of metabolic dysfunction. This study investigates whether incorporating granular diet quality data improves digital screening tools, or if physical measurements render such data redundant.

Methods

We analyzed 20,704 adults from the National Health and Nutrition Examination Survey (1999-2018). We evaluated 13 Healthy Eating Index (HEI-2020) components from 24-hour dietary recalls against a 7-parameter anthropometric model. After Boruta feature selection, we trained Model Averaged Neural Networks and assessed discrimination (ROCAUC) and calibration on a held-out test set.

Results

Boruta identified eight HEI components (e.g., Added Sugars, Dairy) as relevant predictors. The baseline 7-parameter anthropometric model achieved robust discrimination (ROCAUC 0.796, 95% CI 0.785-0.807). Integrating either the HEI Total Score or the eight selected dietary features yielded no significant improvement (ROCAUC 0.796 for both; p > 0.05). Conversely, a diet quality-only model performed poorly (ROCAUC 0.591, 95% CI 0.577-0.605). Calibration remained strong across all models including anthropometrics. However, the diet quality-only model exhibited a limited calibration range, failing entirely to predict probability extremes (<0.273 and >0.818).

Conclusions

While diet quality is biologically critical, its addition via HEI components provides no incremental discrimination or calibration utility to screening models based on robust anthropometrics. The physical phenotype acts as a sufficient downstream integrator of metabolic history. Scalable prediabetes public health screening can primarily rely on simple, objective physical measurements, without requirement for labor-intensive dietary questionnaires.

Key messages

• Detailed diet quality data adds no predictive value to anthropometric diabetes screening tools. Public health screening can rely entirely on simple, scalable physical measurements.

• The physical phenotype acts as a robust downstream integrator of metabolic history. Anthropometric markers fully capture the diabetes risk otherwise encoded by long-term dietary habits.