NHANES accelerometry benchmark — predicting cardiometabolic risk from activity data
New tabular benchmark uses hip-worn accelerometry from 1,381 US adults to predict glycated haemoglobin, triglycerides, and C-reactive protein from activity patterns and lifestyle factors.
TabPFN v2 (a foundation model for tabular data) outperformed ridge regression and XGBoost, achieving R² = 0.156 for HbA1c and R² = 0.383 for CRP.
Dataset reflects real-world survey design — complex sampling, demographic oversampling, subgroup fairness — a gap most clinical ML benchmarks ignore.