Project Summary/Abstract
Wearable devices are the primary method for objectively assessing physical activity (PA) type and energy ex-
penditure (EE) in free-living individuals. Current practice involves using only accelerometer-based devices, which
are generally better for predicting outcomes at the group level rather than the individual level. A ceiling effect
has been reached for accuracy and precision of accelerometer-derived predictions, and thus there is a critical
need for other approaches that can yield more accurate and precise methods to classify PA type and estimate
EE. A potential solution is to combine data from accelerometers with data from other sensors. Accelerometers
record linear acceleration, which captures a large amount of human movement. However, many daily activities
contain turning motions that are not captured by only using accelerometers. Gyroscopes record angular velocity,
and thus may be useful in combination with accelerometers for capturing a richer picture of human movement.
This can result in improved accuracy and precision when assessing PA type and EE. Using an ActiGraph GT9X
(worn on hip, wrists, or ankles), we have previously shown that combining accelerometer and gyroscope data
led to individual-level accuracy improvements of ~6%, compared to accelerometer only. Importantly, this in-
cluded up to 30% improvement for classifying sedentary activities. In addition, classification accuracy between
sedentary and non-sedentary behaviors when using only the accelerometer, ranged from 76.7-96.7% across
wear locations, whereas the gyroscope correctly classified 100% of the time at all wear locations. The overall
objective of this R01 application is to use gold standard measures of EE (doubly-labeled water, room calorimetry
and portable indirect calorimetry) and activity classification (video direct observation) to develop and refine ma-
chine learning algorithms using both accelerometer and gyroscope sensor data. The specific aims of the study
are: 1) Develop and validate gyroscope-inclusive machine learning models that classify PA type and estimate
EE in adults, using a 24-hr stay in a room indirect calorimetry (n=50) and 2-hr of semi-structured activities with
portable calorimetry (n=50); 2a) Assess free-living performance of the models, and 2b) Re-train and refine the
models using free-living data with ground truth from direct observation and portable indirect calorimetry (n = 100
participants during 12 hrs of free-living activity); and 3) Assess validity of EE models during a prolonged free-
living period using the doubly-labeled water technique (n=100). The central hypothesis is that the gyroscope will
provide meaningful and discriminative information on rotational movements that occur during human movement,
thereby complementing the accelerometer data. Combining accelerometer and gyroscope sensor data will im-
prove accuracy and precision for classifying PA type and estimating EE compared to using either sensor alone,
and will have a significant impact on the ability to assess free-living PA in adults.
