Table IV. Techniques to Estimate BR from a Respiratory Signal.

Frequency-based techniques
Spectral analysis: identify the respiratory frequency from a power spectrum calculated using either: the fast Fourier Transform [8] (which can operate on unevenly sampled signals [87]), AR spectral analysis using Burg or Yule-Walker algorithms [199], the Welch periodogram [111], the short-time Fourier transform [192], the Lomb-Scargle periodogram (which can operate on unevenly sampled signals) [52], or sparse signal reconstruction (which can be applied to multiple respiratory signals) [222], [223]. The BR is usually identified as the frequency corresponding to the maximum spectral power in the range of plausible respiratory frequencies although other approaches have been proposed [224]. Identify the respiratory frequency as the dominant frequency of a scalogram calculated using the continuous wavelet transform [41]. Identify the common frequency component in multiple respiratory signals using the weighted multisignal oscillator based least-mean-square algorithm [92]. Estimate instantaneous BR [92] using an adaptive notch filter [106], [173] or an adaptive bandpass filter [155]. Find periodicity using the autocorrelation function [184]. Estimate the instantaneous BR from either a single signal or multiple signals using a bank of notch filters [153], [154]. Autoregressive all-pole modeling, with BR estimated from the frequency of either the highest magnitude pole [10], or the lowest frequency pole [85]. The (order reduced) modified covariance method has also been used [136], [139] Use Gaussian process regression to estimate periodicity [177].
Time-domain breath detection techniques
Detect breaths using peak detection. Detect breaths by identifying zero-crossings with a positive gradient (after detrending) [32]. Detect breaths from peaks and troughs using (adaptive) thresholding to identify those breaths that have been reliably detected [74], [147], [184]