Improve the accuracy and precision of wearable sensor estimates of physical activity and sedentary behavior.
Use of advanced machine learning methods to process and interpret signals from wearable sensors that more accurately estimate the quality and quantity of physical activity and sedentary behavior.
Immediately, improving the activity accuracy of wearable sensors and the activity ‘dose’ recommendations based on them. In the mid-term, better health outcomes by improving consumer wearable sensor activity experiences connected to their broader behaviors, and by personalizing patient activity care plans and monitoring.
Physical activity affects numerous health-related outcomes in children and adults and is increasingly advised as an intervention for preventing or treating many conditions. It is critical to accurately assess the dose of physical activity required for positive health outcomes and actually experienced by an individual. Wearable sensors, accurate over several days or weeks, are the solution to providing such information to the researcher or consumer.
Two strategies are employed to develop validated, innovative methods to process and interpret wearable sensor physical activity and sedentary behavior data. First, the team led by kinesiologist, Patty Freedson, tests the accuracy and precision of wearable monitors by comparing monitor output metric estimates such as steps, energy expenditure, and time spent in moderate to vigorous activity to criterion measures of these variables such as measured energy expenditure and direct observation of behavior. Second, the group led by statistician John Staudenmayer utilizes advanced machine learning methods to process and interpret signal output using artificial neural networks, random forests and other machine learning models to better estimate activity levels. Lab results have shown that these advanced computational methods improve estimates of energy expenditure and time spent in different activity intensity categories compared to the traditional cut-point methods that used simple linear regression modeling.
Working with Enformia as part of an SBIR funded project, we are developing an extensible and scalable commercial software platform that can manage the full lifecycle of objective behavioral measurement data, including measures of physical activity from wearable sensors. We provide feedback on software requirements for data management and guidance on analysis of raw accelerometer physical activity data, advise on data and resultsvisualization techniques and assist in developing representative test datasets. In addition, we are collaborating data analysis software plugin to their platform that leverages our decision tree/neural network algorithms to process free living accelerometer data. Enformia is developing the software platform and the market for it. We will beta test the software on active research projects.
