Breathing rate (BR) is an important physiological indicator that gives information about a variety of chronic diseases. As direct measurements of respiratory devices are uncomfortable for patients, our objective is to obtain an accurate estimation of BR using only PPG signals. To this end, three respiratory modulations are derived from the PPG signals based on amplitude, frequency and baseline wander modulations. These derived modulations are however highly dependent on patient and activity, it is thus difficult to determine the optimal modulation. Therefore, respiratory quality indices (RQI) are introduced to assess the quality of the derived modulations before estimation of BR. These RQIs are based on a set of features (maximum power in the frequency range) extracted from Fourier Transform, autocorrelation and sinusoidal model under the hypothesis that the respiration is a quasi-periodic signal. The selection of the best derived modulations is performed automatically by comparing the RQI scores. This method is compared to other methods in the literature (Pimentel 2016 [1], Karlen2013 [2], Flemming2007 [3], Shelly2006 [4]) using the Capnobase Benchmark dataset. Absolute errors are calculated as the difference between the estimated BR and those derived from the capnometry waveform as a gold standard. For window segments of 32 seconds, Karlen [2]'s method (best reference method) gave a median absolute error (MAE) of 1.2bpm and a 25–75 percentile range in [0.5-3.4] bpm, while our method achieves an MAE of 0.8bpm and a 25- 75 percentile range in [0-4.5]bpm. These results show that the automatic selection using RQI scores is an efficient method for indirect BR estimations based on noisy PPG modulations.
展开▼
