An attention guided lightweight network-based scheme for anxiety detection using multimodal analysis of single-channel wearable ECG and RSP sensor signals

Abstract

This letter presents an attention-guided, lightweight deep learning (DL) network-based approach that utilizes electrocardiogram (ECG) and respiration (RSP) sensor signals to detect various stages of anxiety. For accurate detection, an effective attention mechanism has been incorporated into our proposed DL baseline architecture with a multiobjective loss function. Our proposed model has proven to be highly effective, with minimal trainable parameters and a very simple structural design, achieving an impressive accuracy of 98.67% on a publicly available benchmark dataset in predicting four different anxiety classes. The proposed model has been thoroughly tested using various data window durations, different loss functions, and attention mechanisms. Finally, it has been demonstrated that the proposed architecture, incorporating adaptive attention and a multiobjective loss function, outperforms existing methods in anxiety stages detection.