Researchers built a public benchmark to test how well machine learning can predict cardiometabolic risk from wearable accelerometer data — and the answer is: partially, with asterisks.
The NHANES Accelerometry Cardiometabolic Benchmark draws from a 2003-2006 national health survey, covering 1,381 adults with hip-worn accelerometer readings, blood biomarkers, dietary data, and body measurements. Three modeling approaches — ridge regression, XGBoost, and the tabular foundation model TabPFN v2 — were tested against three targets: glycated haemoglobin (HbA1c, a diabetes marker), fasting triglycerides, and C-reactive protein (a inflammation marker). TabPFN v2 led on two of three, posting R² scores of 0.156 for HbA1c and 0.383 for CRP. Triglycerides were effectively unpredictable (R² below 0.05), which the researchers attribute to strong genetic influence — no lifestyle signal in your step count is going to override that.
The more pointed finding is about uncertainty and fairness. The team applied conformal prediction to generate 90% confidence intervals, and coverage held up for HbA1c and CRP overall — but broke down at the subgroup level. Mexican American participants saw worse HbA1c coverage, a gap the authors flag as clinically meaningful. Aggregate accuracy statistics can mask exactly this kind of disparity.
This is a modest dataset by modern ML standards, and the researchers are upfront about it. But the benchmark fills a real gap: most tabular ML evaluations ignore survey weighting, demographic oversampling, and subgroup equity — the exact complications that show up when you try to deploy these models in a clinic.