When and How to Focus: A Task-Dependent Analysis of Efficient Attentive Architectures for Biosignal Classification

Realizing the vision of Healthcare 5.0 requires diagnostic models that are accurate, yet efficient enough for real-time monitoring on edge devices. Currently, the growing computational cost of large deep learning models continues to create a hurdle to their wide adoption in such resource-constrained settings. Within this context, this paper introduces a lightweight, learnable cross-attention module to efficiently refine fixed self-attentive embeddings from our previously proposed training-free Scattering Transformer (ST). The proposed framework constitutes an efficient attentive architecture that combines the power of the fixed self-attention mechanism of the Scattering Transformer, along with an efficient learnable cross-attention module that enables task-dependent refinement.We conducted an extensive evaluation across benchmark electrocardiogram (PTB- XL, PhysioNet 2017) and heart sound (PhysioNet 2022) datasets to determine the optimal supervised learning architecture for this refinement. Our results reveal that the optimal strategy is critically dependent on the characteristics of the biosignal. For heart sounds, where transient events are crucial, the cross-attentive architecture was essential for high performance. In contrast, for the broad ECG anomaly detection on PTB- XL, the ST’s inherent self-attention representations alone proved most effective. Furthermore, for the specific task of atrial fibrillation detection, our methods achieved promising results that exceeded several state-of-the-art benchmarks. These findings establish a principle of task-dependent architectural restraint, providing a blueprint for sustainable, patient-adaptive biosignal analysis in real-world care-delivery contexts.