Accelerometry-Derived Digital Biomarkers for Cardiometabolic Risk: A Population-Representative Tabular Benchmark with Uncertainty Quantification

Structured tabular data is the dominant format in clinical medicine, yet existing benchmarks fail to reflect key properties of real-world health data: complex survey sampling, demographic oversampling, and fairness requirements. We introduce the NHANES Accelerometry Cardiometabolic Benchmark, derived from the National Health and Nutrition Examination Survey (NHANES) 2003-2006, comprising 1,381 adults with hip-worn accelerometry, fasting laboratory biomarkers, and dietary intake. We evaluate three methods—ridge regression, XGBoost, and the tabular foundation model TabPFN v2—on the prediction of HbA1c, fasting triglycerides, and C-reactive protein (CRP). TabPFN v2 achieves the best overall performance (HbA1c $R^2=0.156$, CRP $R^2=0.383$). We apply split conformal prediction to all models, producing distribution-free 90% prediction intervals, and evaluate demographic coverage equity across sex and race/ethnicity subgroups. Conformal coverage meets or exceeds the 90% target for TabPFN v2 across most subgroups, demonstrating reliable uncertainty quantification in a population-representative setting. All code will be made publicly available.