Model Validation of Physical Activity Energy Expenditure Based on Feature Engineering and Deep Learning

Objective To investigate the effect of wearing positions, inertial measurement unit (IMU) (accelerometers and gyroscopes), and machine learning methods on the prediction of human energy expenditure in walking and running.

Methods The original signal of IMU was preprocessed and a variety of activity features were extracted. The machine learning method based on feature engineering and the convolutional neural network method based on deep learning were modeled by cross-validation.

Results In walking and running activities, the performance of ankle model was better than that of hip model for different wearing parts of IMU. For different types of motion signals, the performance of accelerometer model was better than that of gyroscope model. For different machine learning models, machine learning methods based on feature engineering were generally superior to convolutional neural network methods based on deep learning.

Conclusion The artificial neural network model based on multi-signal fusion can generalize and stabilize better than the other models under different exercise intensities, which is recommended to estimate energy expenditure for walking and running activities.