Self-Supervised Machine Learning to Characterize Step Counts from Wrist-Worn Accelerometers in the UK Biobank.

Area of Science:

• Biomedical Engineering
• Epidemiology
• Machine Learning

Background:

• Accurate physical activity measurement is crucial for health research.
• Wrist-worn devices often have >20% error in step counting during free-living conditions.
• Developing reliable step count algorithms is essential for large-scale studies.

Purpose of the Study:

• To develop and validate a machine learning-based step counting algorithm for wrist-worn accelerometers.
• To assess the association between objectively measured daily step count and mortality risk.
• To improve the accuracy of physical activity quantification in large cohorts.

Main Methods:

• A self-supervised machine learning model was trained on annotated free-living step data.
• External validation was performed using an independent open-source dataset.
• Cox regression analysis assessed the association between step count and mortality in 75,263 UK Biobank participants.

Main Results:

• The developed algorithm achieved a mean absolute percent error of 12.5%, outperforming reference models (65%-231%).
• Daily step counts of 6430-8277 were associated with a 37% lower risk of fatal cardiovascular disease.
• These step counts also correlated with a 28% lower risk of all-cause mortality over 7 years.

Conclusions:

• An open and transparent method significantly enhances step count accuracy from wrist-worn accelerometers.
• The findings demonstrate a clear link between higher step counts and reduced mortality risk.
• This work supports public health initiatives promoting physical activity and may inform future guidelines.