Applying a novel statistical pipeline — combining accelerated failure time models, spline regression, and simulation-extrapolation (SIMEX) — to UK Biobank accelerometer data reveals that standard analyses systematically underestimate physical activity's protective effect against cardiovascular disease. After correcting for machine-learning classification errors inherent in wrist-worn accelerometer data, the time ratio comparing the 95th versus 5th percentile of total daily activity rose from 1.38 to 1.56, a meaningful upward revision. Notably, the corrected dose-response curve was also nonlinear — a shape the uncorrected model failed to capture entirely. Sex-stratified sensitivity analyses suggest women derive substantially larger corrected benefits than men.
This finding matters because the entire physical activity epidemiology literature may be systematically underestimating activity-health associations due to measurement error in accelerometer-derived metrics — a bias that compounds when machine learning classifiers introduce their own misclassification noise. If confirmed, recalibrated dose-response curves could shift public health guidelines toward stronger, more precisely shaped activity recommendations. The SIMEX methodology itself is not new but has rarely been applied to nonlinear survival models with wearable data, making this a genuinely methodological contribution with broad applicability. Limitations include the observational UK Biobank design (no causal inference), potential residual confounding, and the reliance on simulation assumptions to characterize measurement error structure. The sex difference finding, while intriguing, needs replication. As a preprint posted on medRxiv and not yet peer-reviewed, these results — and the corrected effect estimates — should be interpreted cautiously until independent validation occurs.
