Multi-frequency Model Fusion for Robust Breathing Rate Estimation

GE Di1, guy carrault2, soumaya khreis3
1LTSI, 2LTSI IINSERM 1099 UR1, 3PHD student


In this study, we aim to obtain an accurate estimation of BR using indirectly monitored signals, such as PPG or ECG. We have previously proposed signal quality based selection of derived modulations and a Kalman smoothing fusion strategy. We further investigate here the pertinence of enhancing model complexity by introducing multiple frequency dynamics. Performances are compared to reference methods (Pimentel2016, Karlen2013) on the Capnobase Benchmark dataset (

The BR estimation by fusion is illustrated through an example of ECG derivation inputs such as R-wave peak amplitude (RPA), Q-R wave amplitude (QRA), R-R temporal series (RSA) and the triangle area of Q,R,S coordinates (AQRS). The application with PPG signals on the otherhand would involve other modulations (RIIV, RIAV and RIFV) as in the previous studies of Pimentel and Karlen. Respiration quality indexes (RQI) are used in order to keep only sinusoid-like modulations for fusion. Metrics such as the energy ratio of dominant frequency components and auto-correlations (see our previous work on the RQI in cinc) are typically used. The originality of this study is to consider multi-frequency model in the Kalman fusion step as opposed to common single frequency models. The strategy is justified from the observation that spectral energies of respiration modulations often include multiple (high-)frequency modes especially for square-formed modulations.

The multiple-frequency dynamics within the Kalman fusion better fit the respiration modulations and thus yield significant performance gains compared to state of the art reference methods (0.3 vs 0.9 in median bpm error for PPG and 0.1 vs 1.1 for ECG). While Karlen’s approach also rejected around 36% temporal segments for estimation, we reported estimation results on the whole Capnobase and applied the fusion method on available ECG signals, never exploited to our knowledge.