An evaluation of ECG data fusion algorithms for wearable IoT sensors

• Published in: Information fusion Vol. 96; pp. 237 - 251
• Main Authors: John, Arlene, Padinjarathala, Antony, Doheny, Emer, Cardiff, Barry, John, Deepu
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2023
• Subjects: Bayesian filtering; Data fusion; Electrocardiography; Heart rate estimation; Kalman fusion; Signal quality indicators; Electrocardiography; Data fusion; Signal quality indicators; Bayesian filtering; Kalman fusion; Heart rate estimation
• ISSN: 1566-2535; 1872-6305; 1872-6305
• DOI: 10.1016/j.inffus.2023.03.017

In wearable sensing, accurate estimation of physiological parameters is paramount, although these signals can be corrupted by noise. The fusion of data from multiple sensor sources has the potential to enhance accuracy, even in the presence of disruptive noise. This paper aims to introduce and compare various existing state-of-the-art and novel data fusion techniques to improve the reliability of heart rate estimation. The comparisons were implemented using the MIT-BIH Arrhythmia database with additive noise signals taken from MIT Noise Stress Test Database. When it comes to the challenging low signal-to-noise ratio (SNR) regions, the Kalman fusion and the α-trim mean filtering approach exhibits the best performance. The Kalman fusion approach dominates when both channels are corrupted, while the α-trim mean filtering elimination algorithm takes the lead when at least one channel is clean. To make the most of these strengths, we have developed an innovative algorithm that can switch between the two fusion methods based on a signal quality indicator (SQI) that serves as a surrogate SNR. This algorithm outperforms the baseline 2-channel RR-interval averaging approach by ≃54% and ≃21% at SNRs of 20dB and −20dB respectively. Moreover, it outperforms other cutting-edge heart rate estimation methods. •Investigation of fusion methods for heart rate estimation from ECG signals.•Signal quality aware fusion algorithm selection.•Fusion methods analyzed at varying input signal-to-noise ratios.•Signal quality index proposed as a surrogate for input signal-to-noise ratio.•Fusion selection method proposed for varying input signal-to-noise ratios.