Estimating within-stride metabolic cost from stride-average data using autoencoders and expander networks.

Area of Science:

• Biomechanics
• Computational modeling
• Machine learning

Background:

• Aging increases walking oxygen consumption by over 30%, compromising mobility and leading to physical degeneration.
• The underlying reasons for increased metabolic cost during gait are not fully understood.
• Current motion capture systems cannot directly measure metabolic cost fluctuations within the gait cycle.

Purpose of the Study:

• To compute the metabolic cost time series from mean values using neural network approaches.
• To investigate the effectiveness of autoencoders (AEs) and expanders for this task.
• To address the challenge of characterizing within-stride metabolic cost fluctuations.

Main Methods:

• Two neural network models, autoencoders (AEs) and expanders, were developed to map mean metabolic cost to time series.
• Networks were trained using a computational walking model simulating gait under 35 robotic perturbations.
• Model validation was performed using estimated metabolic costs for unperturbed gait cycles.

Main Results:

• Expanders and AEs without tied weights performed best with nonlinear activation functions.
• AEs with tied weights performed best with linear activation functions.
• Expanders unexpectedly outperformed AEs in estimating metabolic cost time series.

Conclusions:

• Neural network approaches, particularly expanders, can effectively compute metabolic cost time series from mean values.
• This method offers a potential solution for estimating within-stride metabolic cost fluctuations.
• Future work will focus on overcoming reliance on time series for initial training.