Egocentric 3D Skeleton Learning in a Deep Neural Network Encodes Obese-like Motion Representations

Area of Science:

• Biomedical Engineering
• Computational Biology
• Deep Learning Applications

Background:

• Obesity is a significant global health issue driven by poor diet.
• Accurate dietary tracking for obese individuals remains a challenge.
• 3D motion capture's potential for early obesity detection is underexplored.

Purpose of the Study:

• To explore the use of deep learning and 3D motion capture for detecting obesity-related motion patterns.
• To evaluate different deep recurrent networks for analyzing skeletal data.
• To identify optimal methods for encoding obese-like motion representations.

Main Methods:

• Utilized deep LSTM networks trained on individual identity (identity-trained deep LSTM network).
• Analyzed 3D time-series skeleton data from diet-induced obese mouse models.
• Compared allocentric and egocentric viewpoints and various recurrent networks (RNN, GRU, LSTM).
• Employed a support vector classifier with latent features to test motion representations.

Main Results:

• The identity-trained deep LSTM network achieved optimal performance.
• An egocentric viewpoint combined with the identity-trained deep LSTM network yielded the best results.
• The models effectively encoded obese-like motion representations.

Conclusions:

• Deep learning, specifically identity-trained deep LSTM networks with egocentric viewpoints, can effectively detect obesity-related motion in mouse models.
• This approach offers a novel method for identifying health risks and early signs of obesity.
• The findings suggest potential applications for human obesity detection.