Deep Learning-Based Obesity Identification System for Young Adults Using Smartphone Inertial Measurements

Area of Science:

• Biomedical Engineering
• Computer Science
• Public Health

Background:

• Adolescent obesity is a significant public health issue requiring innovative detection methods.
• Current methods for obesity identification may lack accessibility or scalability.
• Smartphone technology offers a potential platform for widespread health monitoring.

Purpose of the Study:

• To develop and evaluate a deep learning framework for identifying adolescent obesity using smartphone inertial measurements.
• To compare the performance of Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and a hybrid CNN-LSTM model for obesity recognition.
• To assess the feasibility of using mobile health applications for gait-based obesity screening.

Main Methods:

• Collected smartphone inertial data (accelerometer, gyroscope, rotation vector) from 138 subjects via a mobile health app.
• Analyzed gait patterns to extract features indicative of obesity.
• Trained and tested three deep learning models: CNN, LSTM, and a hybrid CNN-LSTM architecture.
• Evaluated model performance based on accuracy in distinguishing between normal-weight and obese adolescents.

Main Results:

• The hybrid CNN-LSTM model demonstrated the highest accuracy at 97% for obesity identification.
• The LSTM model achieved 96.31% accuracy, and the CNN model achieved 95.81% accuracy.
• The framework successfully utilized gait analysis from smartphone sensors for obesity detection in adolescents.

Conclusions:

• Deep learning models, particularly the hybrid CNN-LSTM, show high efficacy in identifying adolescent obesity using smartphone gait data.
• Smartphone-based gait analysis presents a viable, non-invasive approach for obesity screening and early intervention.
• Future research should focus on expanding the model's generalizability to diverse gait patterns and clinical populations.