A Robust PPG Heart Rate Estimation Network Based on Frequency-Domain NLMS and Dual-Stream Hierarchical Self-Attention

Abstract

The heart rate estimation based on the wearables' photoplethysmography (PPG) sensor always faces serious motion artefacts (MA) under high-intensity exercise. Although adaptive filtering and deep learning methods have demonstrated effectiveness in practice so far; currently, there are problems such as system divergence under the influence of strong power signals, poor generalization due to insufficiently trained high-heart-rate sample sets. The main deficiencies in the existing systems are as follows: Filter instability due to significant noise; The coupling of physiological and motion variables; Imbalanced datasets. To solve the above-mentioned problems, a full-ended design of frequency-domain signal processing and deep neural networks is proposed. Through an energy-normalized NLMS adaptive filter to establish the system's absolute stability constraint condition, in conjunction with a dual-stream multiple-head self-attention (MHSA) mechanism for controllably managing features weights; Moreover, by incorporating a specific data augmentation strategy. On the Public PPG-DaLiA dataset, this built network achieves an average absolute error of 1.81 beats per minute (BPM). Based on the results, a feasible approach can be used in clinical practice to improve the credibility of wearable device applications by providing reasonable uncertainly ranges.