Wrist-independent energy expenditure prediction models from raw accelerometer data

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Human Movement Analysis

Background:

• Accurate energy expenditure (EE) estimation is crucial for health monitoring and research.
• Wrist-worn accelerometers are promising for unobtrusive EE tracking.
• Developing robust models that account for accelerometer placement is essential.

Purpose of the Study:

• To develop artificial neural network (ANN) models for predicting EE from wrist accelerometer data.
• To develop ANNs capable of detecting the wrist on which the accelerometer is worn.
• To evaluate the accuracy of EE prediction models using same-wrist versus opposite-wrist data.

Main Methods:

• Forty-four adults participated in a 90-minute activity protocol involving 14 different activities.
• GENEActiv accelerometers were worn on both wrists, and EE was measured using a portable metabolic analyzer.
• ANNs were developed using leave-one-out cross-validation, testing various feature sets (raw, vector magnitude, absolute values) and training/testing wrist combinations.

Main Results:

• ANNs using raw data from the same wrist achieved high EE prediction accuracy (r=0.84, RMSE=1.25-1.26 METs).
• Opposite-wrist prediction accuracy was lower (r=0.60-0.64, RMSE=1.93-2.01 METs).
• Preprocessing data into absolute values resulted in high, wrist-independent EE prediction accuracy (r=0.80-0.83, RMSE=1.30-1.49 METs).
• Wrist detection ANNs achieved 100% accuracy in identifying accelerometer placement.

Conclusions:

• Highly accurate, wrist-independent EE prediction models can be developed by computing absolute values of raw acceleration data prior to ANN development.
• This preprocessing method offers a significant advancement for improving the predictive accuracy of wrist-worn accelerometers.
• The ability to accurately detect wrist placement further enhances the utility of these devices.