Respiratory rate estimation from the ECG using an instantaneous frequency tracking algorithm

Highlights

• Two respiration-related waveforms are derived from the ECG.

• A frequency tracking algorithm estimates the respiratory rate from them in combination.

• This combination is more effective than using either waveform separately.

• The algorithm is automatic and real-time.

Abstract

Monitoring the respiratory rate (RR) is important in many clinical and non-clinical situations but it is difficult in practice, for existing devices are obtrusive, bulky and expensive. The extraction of the RR from the routinely acquired electrocardiogram (ECG) has been proposed lately. Two approaches exist, one exploiting the modulation of the heart rate by the respiration, known as the respiratory sinus arrhythmia (RSA) and the other using the modulation by the respiration of the R-peak amplitudes (RPA). In this study, the weighted multi-signal oscillator based band pass filtering (W-OSC) algorithm is applied to track the common frequency in the RSA and RPA waveforms simultaneously, as an estimate of the instantaneous RR. On the public PhysioNet Fantasia data set, it is shown that the presented method is automatic, instantaneous and comparable in accuracy to the state-of-the-art.

Introduction

The respiratory rate (RR) is one of the human vital signs that need to be monitored in clinical and non-clinical applications for diagnosis and control purposes [1]. It is currently difficult to accurately and continuously monitor the RR, for the apparatus and devices are intrusive, expensive and uncomfortable for the patient [2]. It is therefore of great interest to provide easy and inexpensive means for the accurate, continuous and convenient monitoring of the RR. The respiratory and heart activities are linked through physiological processes. The respiration modulates the heart rate such that it increases during inspiration and decreases during expiration [3]. A waveform can be extracted from the heart rate time series representing this modulation, which is referred to as respiratory sinus arrhythmia (RSA). Furthermore, the filling and emptying of the lungs during respiration causes a rotation of the electrical axis of the heart and a change in the impedance of the thorax, which yields changes in the electrocardiogram (ECG) beat morphology. As a result, the R-peak amplitudes are modulated by the respiratory activity. A waveform can be extracted from the ECG representing this modulation, which is referred to as the R-peak amplitude (RPA). In the last twenty years, many researchers have investigated the possibility of deriving the RR by exploiting the influence of respiration on the heart rate or ECG beat morphology by using the RSA or RPA waveforms. A summary of studies prior to 2007 can be found in [3].

In a more recent work, a windowed spectral analysis was applied to extract the RR from the RSA waveform [4]. A windowed temporal analysis using peak count was also applied to estimate the RR from the RSA waveform [4], [5] as well as from the RPA waveform [5]. Furthermore, correlation analysis was used to compute the RR from the RSA waveform [4]. Past studies have highlighted the difficulty of deriving, for comparison purposes, a reference RR from a respiratory signal recorded by nasal thermistry or pneumography, as the latter is neither band-pass, nor stationary [4]. Proposed solutions consisted in applying the same frequency estimation methods as for the RSA or RPA waveforms to the respiratory signal to estimate the reference RR.

In a pioneering work, Orphanidou et al. fused spectral information form ECG-derived RSA and RPA waveforms in order to derive the RR [6]. The most dominant peak from the autoregressive (AR)-estimated spectra of the RSA and RPA waveforms was selected according to several criteria as representing the RR. Vehkaoja et al. also used RSA and RPA signals simultaneously and combined their temporal maxima and minima counts after ad hoc filtering to estimate the RR [7]. However, none of these two methods provides a robust, real-time and automatic means to estimate the RR continuously. Both discarded parts of records and required special subject-dependent treatment.

Motivated by the recent contribution of Orphanidou et al. to fuse respiratory information from the RSA and RPA waveforms extracted from the ECG [6], the purpose of the present study was to estimate the instantaneous real-time RR by using the weighted multi-signal oscillator based band-pass filtering (W-OSC) algorithm [8] to track the common frequency component present in the RSA and RPA waveforms. This multi-signal frequency tracking method operates recursively on several signals simultaneously to track a common frequency component. This method has been shown to successfully track a common frequency component in biomedical signals such as electroencephalogram signals [9]. Furthermore, it is instantaneous and provides an automatic approach to RR estimation from the RSA and RPA, in contrast to the ad hoc processing proposed in [6]. Therefore it can be implemented in a real-time setting to estimate the RR continuously without the need for subject-dependent adjustment. The performance of this tracking method was assessed on the single-lead ECG recordings of the PhysioNet Fantasia data set, using a reference RR computed from the combination of eight different frequency estimates of the simultaneously recorded respiratory signal.