New Computer Model Improves Tracking of Physical Activity From Wrist Sensors
Researchers built and tested a new computer model called ActiNet to classify physical activity intensity using wrist-worn sensor data. The new model performed better than a standard baseline method across different age and sex groups.
In short: Researchers built and tested a new computer model called ActiNet to classify physical activity intensity using wrist-worn sensor data. The new model performed better than a standard baseline method across different age and sex groups.
Fitness trackers on our wrists collect massive amounts of movement data, but turning those numbers into clear health insights is a major challenge for scientists.
What happened, in plain words
Researchers led by authors including A. Acquah, S. Chan, and A. Doherty studied data from 151 participants in the CAPTURE-24 project, ranging in age from 18 to 91 years old, with a mean age of 42 and 66 percent female. They trained a model called ActiNet, which uses a self-supervised deep learning network paired with hidden Markov model smoothing. They compared ActiNet against a baseline random forest method and found that ActiNet achieved higher performance scores for identifying activity intensity.
Key points
- Study participants and data: The research used up to 24 hours of wrist-worn accelerometer data collected from 151 participants in the CAPTURE-24 group.
- The ActiNet model: ActiNet consists of HARNet, an 18-layer modified ResNet-V2 model that uses self-supervised learning, followed by hidden Markov model smoothing.
- Performance comparison: ActiNet reached a mean macro F1 score of 0.82 and a mean Cohen's kappa score of 0.86, beating the baseline random forest model scores of 0.76 and 0.80.
- Consistent results: The performance improvements of the new model stayed consistent across different subgroups based on age and sex.
Terms explained
- accelerometer — A sensor that measures movement, acceleration, and physical motion. Example: The step-tracking device worn on your wrist that counts your daily motion.
- self-supervised learning — A type of machine learning where a computer model teaches itself to understand patterns from unlabeled data before doing specific tasks. Example: A student looking at hundreds of unsorted puzzle pieces on a table to figure out how they fit together before trying to solve the full picture.
- hidden Markov model — A mathematical tool used to predict a sequence of events based on changing conditions. Example: Guessing whether it is raining outside throughout the day just by checking whether people carry umbrellas into a building.
Why it matters
Accurate tools help researchers study how physical activity relates to health outcomes in large-scale epidemiological studies, which investigate health patterns in human populations.
What we still don't know
The extent to which these models make tangible improvements to classification performance and the effect on predicted daily activity intensity compositions was previously unknown. The study relies on the CAPTURE-24 population, and the authors note that these findings encourage future use but do not represent a final solution for all possible populations or medical contexts.
Source: Scientific Reports. The original is licensed CC BY. This text is an AI-assisted adaptation (summarized, simplified and translated) and may differ from the original.