TabPFN v2 beats XGBoost on NHANES cardiometabolic benchmark from wearables

Structured tabular data is the backbone of clinical medicine, but existing benchmarks ignore real-world complexities like survey sampling and subgroup fairness. In a new paper accepted at the SD4H Workshop at ICML 2026, Federico Felizzi presents the NHANES Accelerometry Cardiometabolic Benchmark, derived from NHANES 2003–2006. It includes 1,381 adults with hip-worn accelerometry, fasting biomarkers, dietary intake, and anthropometrics. The benchmark evaluates three tabular learning methods—ridge regression, XGBoost, and the foundation model TabPFN v2—to predict glycated haemoglobin (HbA1c), C-reactive protein (CRP), and fasting triglycerides from activity phenotypes.

TabPFN v2 leads overall, with HbA1c R²=0.156 and CRP R²=0.383, while triglycerides remain largely unpredictable (R²<0.05), consistent with known genetic dominance. The paper also applies split conformal prediction to generate distribution-free 90% prediction intervals, then evaluates coverage equity across sex and race/ethnicity subgroups. Marginal coverage meets the target for CRP and HbA1c but falls short for triglycerides. At the subgroup level, localized undercoverage emerges—e.g., HbA1c predictions for Mexican American participants—illustrating the gap between marginal guarantees and fair conditional coverage for clinical use. Code and data are publicly available.