XNet: Enhancing Physical Activity Intensity Assessment With Attentional Multidomain Fusion and Visual Analytics.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Public Health

Background:

• Sedentary behavior (SB) is a significant global health issue requiring precise physical activity (PA) intensity monitoring.
• Conventional machine learning (ML) models face challenges in generalizing across diverse populations, sensors, and activities, impacting real-world accuracy.

Purpose of the Study:

• To introduce XNet, a dual-domain deep learning (DL) model designed for accurate PA intensity classification and energy expenditure estimation.
• To enhance the generalization and robustness of PA monitoring systems.

Main Methods:

• Developed a hierarchical multihead DL architecture (XNet) that extracts temporal and frequency features from multiple sensors.
• Implemented a two-stage attentional feature fusion (AFF) module for integrating sensor and domain-specific features.
• Validated XNet on multiple public datasets and a new dataset comprising 105 participants.

Main Results:

• XNet achieved a 70.5 F1-score in cross-dataset evaluations and 77.0 in open-set scenarios, outperforming existing DL and ML baselines.
• Demonstrated robust sedentary detection with an 88% true positive rate (TPR).
• Showcased that lightweight 1D-convolutional spectral encoders offer superior out-of-distribution generalization compared to transformers and GAT networks.

Conclusions:

• XNet's hierarchical approach and AFF module provide superior accuracy, efficiency, and robustness for PA intensity monitoring.
• The model's adaptability to physiological signals and low inference latency (~25 ms) support on-device deployment.
• Interpretable attention weights and a visual analytics framework promote transparency and expert auditing in health monitoring.