Improving energy expenditure estimates from wearable devices: A machine learning approach.

Area of Science:

• Bioenergetics
• Wearable technology
• Machine learning

Background:

• Quantifying continuous, free-living energy expenditure (EE) is crucial for advancing bioenergetics research.
• Current methods may not fully capture the nuances of EE during daily life.
• Wearable devices offer a promising avenue for continuous physiological data collection.

Purpose of the Study:

• To apply a non-linear, machine learning algorithm (random forest) to predict minute-level EE.
• To investigate the predictive accuracy of various wearable devices and physiological signals.
• To compare machine learning predictions against indirect calorimetry as a criterion measure.

Main Methods:

• Utilized acceleration, heart rate, body temperature, and galvanic skin response data from Fitbit Charge 2, Polar H7, SenseWear Armband Mini, and Actigraph GT3-x.
• Employed a leave-one-out cross-validation approach with 59 subjects across sedentary, ambulatory, household, and cycling activities.
• Compared model predictions to indirect calorimetry (Vyntus CPX) for accuracy assessment.

Main Results:

• Achieved correlations of at least r = 0.85 across all activities.
• Reported root mean squared error ranging from 1 to 1.37 METs.
• Demonstrated statistically equivalent performance to the criterion measure, with lower error for Actigraph and Sensewear models compared to manufacturer estimates.

Conclusions:

• Non-linear machine learning models applied to wearable devices achieve high accuracy in estimating EE.
• This approach provides a viable means to quantify the energy cost of free-living activities.
• Wearable technology integrated with advanced algorithms can significantly enhance bioenergetics studies.