Accelerometry-derived digital biomarkers are transforming cardiometabolic risk assessment, according to a new benchmark study by Federico Felizzi. Published on June 29, 2026, this research utilizes data from 1,381 adults participating in the NHANES 2003-2006 survey, highlighting the potential of these digital tools in clinical settings.
Understanding the NHANES Accelerometry Benchmark
The NHANES Accelerometry Cardiometabolic Benchmark addresses the limitations of existing clinical benchmarks by reflecting complex survey sampling and demographic factors. Felizzi's study introduces structured tabular data that includes fasting laboratory biomarkers, dietary intake, and anthropometrics, making it a comprehensive resource for researchers and healthcare professionals.
With this benchmark, the research evaluates three tabular learning methods: ridge regression, XGBoost, and the advanced model TabPFN v2. The results show that TabPFN v2 outperforms the others in predicting glycated hemoglobin (HbA1c) and C-reactive protein (CRP), achieving an R² of 0.156 for HbA1c and 0.383 for CRP. However, the study notes that triglycerides remain largely unpredictable, with an R² of less than 0.05, reaffirming genetic dominance in this area.
Implications for Clinical Fairness
This benchmark not only aims to improve prediction accuracy but also evaluates demographic coverage equity across various sex and race/ethnicity groups. The study applies split conformal prediction to generate distribution-free 90% prediction intervals. While marginal coverage meets the 90% target for CRP and HbA1c, it falls short for triglycerides.
