Transforming physical fitness and exercise behaviors in adolescent health using a life log sharing model

This study investigates the potential of a deep learning-based Life Log Sharing Model (LLSM) to enhance adolescent physical fitness and exercise behaviors through personalized public health interventions.

We developed a hybrid Temporal-Spatial Convolutional Neural Network-Bidirectional Long Short-Term Memory (TS-CNN-BiLSTM) model. This model integrates temporal, textual, and visual features from multimodal life log data (exercise type, duration, intensity) to classify and predict physical activity behaviors. Two datasets, Geo-Life Log (with location data) and Time-Life Log (without location data), were constructed to evaluate the impact of spatial information on classification performance. The model utilizes CNNs for local feature extraction and BiLSTM networks to capture temporal dynamics, maintaining user privacy.

The TS-CNN-BiLSTM model achieved an average classification accuracy of 99.6% across eight physical activity types, outperforming state-of-the-art methods by 1.9-4.4%. Temporal features were identified as crucial for detecting recurring behavioral trends and periodic exercise patterns.

These findings demonstrate the efficacy of integrating multimodal life log data with deep learning for accurate physical activity classification. The high accuracy of the TS-CNN-BiLSTM model supports its potential for developing personalized health promotion strategies, including tailored interventions, behavioral incentives, and social support mechanisms, to enhance adolescent engagement in physical activities and advance public health education and personalized health management.