Machine Learning To Predict Energy Expenditure And Type Of Physical Activity From Accelerometer And Heart Rate Data

Purpose Wrist accelerometers are being used in population level surveillance (i.e. NHANES) of physical activity (PA) but more research is needed to evaluate the validity of a wrist-worn device for predicting PA. In this study we compare accelerometers worn on the wrist and each hip for predicting PA type and energy expenditure (EE) using machine learning algorithms. We also investigate the added value of including heart rate (HR) data in making predictions.

Methods Forty adults (21 women, 19 men; mean age = 35.8 ±12.1 yrs; BMI = 24.8 ± 2.9) performed 8 locomotion and household activities for 6 minutes in a lab setting. Participants wore three ActiGraph GT3X+ accelerometers (left hip, right hip, non-dominant wrist), a HR monitor (Polar RS400), and a portable indirect calorimeter (COSMED K4b2). METs were computed for each minute of ventilatory data and 154 features were extracted from each minute of accelerometer data. HR (beats per minute) was used as an additional feature. We developed two different predictive models: a random forest classifier to predict activity type from these features and a random forest of regression trees to estimate EE. Predictions were evaluated using leave-one-user-out cross-validation.

Results For predicting four activity types (household, stairs, walking, running), the hip accelerometer obtained 88.9% accuracy while the wrist accelerometer obtained 82.0% accuracy. Combining data from multiple accelerometers or including HR did not significantly improve these results. In predicting all 8 activities (laundry, window washing, dusting, dishes, sweeping, stairs, walking, running), the (left) hip and wrist accelerometers alone obtained 69.0% and 74.2% accuracy, respectively. Combining hip and wrist data led to 80.0% accuracy, but adding HR did not significantly improve results. Predicting METs using the (left) hip or wrist devices alone obtained root mean square errors (rMSE) of 1.138 and 1.249, respectively. Including HR data and multiple accelerometers improved MET estimation (rMSE = 1.061 with combined wrist, left hip and HR). There was no significant bias.

Conclusions Results demonstrate the validity of random classification and regression forests for activity type and MET prediction using accelerometers. The wrist accelerometer was more useful in predicting activities with significant arm movement (e.g., household activities), while the hip accelerometer was superior for predicting locomotion and estimating EE. Results also demonstrate that HR data does not significantly improve activity classification but does improve MET estimation.