Personalized Energy Expenditure Estimation: Visual Sensing Approach With Deep Learning.

Area of Science:

• Biomedical Engineering
• Computer Science
• Human Physiology

Background:

• Calorimetry is the gold standard for measuring energy expenditure but is costly and intrusive.
• Personalized energy expenditure models are crucial as individuals vary in metabolic efficiency.
• Few-shot learning advances enable personalized deep learning models without extensive data.

Purpose of the Study:

• Identify optimal daily living activities for calibrating a vision-based personalized deep learning calorie estimation system.
• Evaluate the efficacy of short activity durations for model calibration.
• Compare calibration performance using limited actions versus longer, varied activity sequences.

Main Methods:

• Utilized the SPHERE Calorie dataset (10 participants, 11 activities, 4.5 hours).
• Employed deep learning for calorie prediction from video data, regressing predictions against calorimeter ground truth.
• Personalized models using brief video segments (32 seconds) of specific activities.

Main Results:

• A single activity, 'wipe', achieved 1.40 Mean Squared Error (MSE).
• A pair of activities, 'sweep' and 'sit', yielded the best calibration with 1.09 MSE.
• Calibration using 'sweep' and 'sit' (1.09 MSE) closely matched using a 30-minute sequence (1.06 MSE).

Conclusions:

• A vision-based deep learning system for energy expenditure estimation can be effectively calibrated.
• Minimal data (32 seconds of 'sweep' and 32 seconds of 'sit') is sufficient for personalizing the system.
• This approach offers a feasible alternative to traditional calorimetry for daily energy expenditure monitoring.