Deep Attention-Based Signal Analytics for Automated Interpretation of High-Dimensional Biomedical Signals

Abstract

The recent accelerated biomedical sensing technology has resulted in the creation of high
dimensional, complex, and noisy physiology signals posing an enormous challenge to reliable
automated interpretation of such physiology signals with traditional signal processing and
learning methods. The currently used techniques have been particularly poor at capturing
long-range correlations, inter-channel interactions, and clinically important discriminative
structures that are inherent to such large-dimensional biomedical data, which limits their
analysis fidelity and predictive ability. In order to deal with these shortcomings, the present
paper suggests a signal analytics framework based on deep attention to integrate high-level
time-frequency signal representation with attention-driven deep architecture to facilitate
strong and automated interpretation of biomedical signals. The suggested framework
builds on the idea of attention-enhanced neural architecture and transformer-based and
relies on the selective emphasis of informational temporal and spectral features and the
muting of redundant and noise-sensitive elements. Representative biomedical signal datasets
are experimentally assessed with the standard metrics of classification, namely accuracy,
precision, recall, F1-score, specificity and AUCROC, and compared to traditional machine
learning and deep learning baselines. The findings reveal statistically significant progress in
the performance of the classification, superior interpretability by visualising attention, and
competitive computational efficiency, and further show that the proposed methodology
is a viable way of achieving scalable and reliable biomedical signal analytics in the next
generation intelligent healthcare systems.