Accelerometry Predicts Cardiometabolic Risk in New Benchmark

A new benchmark dataset, the NHANES Accelerometry Cardiometabolic Benchmark, has been developed to advance the use of digital biomarkers in healthcare. This dataset, derived from the NHANES 2003-2006 study, includes accelerometry data from hip-worn devices, alongside fasting laboratory biomarkers, dietary intake, and anthropometrics for over a thousand adults. Its design specifically addresses real-world challenges in clinical data, such as complex survey sampling and demographic oversampling, which are often overlooked in existing benchmarks. The study evaluated three tabular learning methods—ridge regression, XGBoost, and TabPFN v2—for their ability to predict key cardiometabolic indicators like glycated haemoglobin (HbA1c), fasting triglycerides, and C-reactive protein (CRP). TabPFN v2 emerged as the top performer for HbA1c and CRP prediction, though triglycerides remained largely unpredictable, consistent with their known genetic dominance. Crucially, the research also applied split conformal prediction to quantify uncertainty and assess fairness across demographic subgroups. While overall prediction intervals met targets for CRP and HbA1c, localized undercoverage was observed for specific subgroups, such as Mexican American participants for HbA1c. This highlights a critical gap between marginal guarantees and the conditional coverage needed for equitable clinical application.