E-BDL: Enhanced Band-Dependent Learning Framework for Augmented Radar Sensing

Area of Science:

• Radar sensing technology
• Deep learning applications in healthcare
• Signal processing for physiological monitoring

Background:

• Radar sensors offer nonintrusive kinetic and physiological motion capture, preserving privacy.
• Deep learning (DL) enhances radar sensing for gait recognition and vital-sign measurement.
• Band-dependent patterns in time-frequency representations (TFRs) challenge DL models by potentially obscuring low-power, frequency-specific features.

Purpose of the Study:

• To propose an Enhanced Band-Dependent Learning (E-BDL) framework to address challenges in radar sensing using DL.
• To improve the detection and utilization of band-dependent features within sub-frequency bands for enhanced classification.
• To enhance the performance and interpretability of DL-based radar sensing models in healthcare applications.

Main Methods:

• Developed an E-BDL framework with adaptive sub-band filtering, representation learning, and sub-view contrastive modules.
• Utilized TFRs to analyze band-dependent features in radar signals.
• Validated the framework on datasets for Alzheimer's disease (AD) and AD-related dementia (ADRD) risk evaluation, and hemodynamics scenario classification.

Main Results:

• E-BDL-ResNet demonstrated competitive performance in hemodynamics scenario classification compared to recent methods.
• E-BDL-ResNet achieved superior performance in ADRD risk evaluation across all candidate models.
• The framework effectively identified salient sub-bands in TFRs, improving DL model performance and interpretability.

Conclusions:

• The E-BDL framework enhances representation learning by detecting critical sub-bands in TFRs.
• E-BDL shows significant potential as a clinical tool for gait abnormality recognition and vital-sign monitoring.
• The proposed method improves both the performance and interpretability of deep learning models in radar sensing for healthcare.