Biomedical signals are often rhythmical and their morphologies change slowly over time. Arterial blood pressure and electrocardiogram signals are good examples with such property. It is of great interest to extract clinically useful information such as the instantaneous frequency (i.e. heart rate) and morphological changes (e.g. pulse pressure variation) from these signals. Conventional filtering methods such as the Kalman filter are not suitable for estimating the instantaneous frequency of quasiperiodic signals due to the non-Gaussian multi-modal property of its posterior distribution. One possible alternative is particle filters that are increasingly used for nonlinear systems and non-Gaussian posterior state distributions. However, canonical particle filters suffer from the problems of sample degeneracy and sample impoverishment and are not well suited to non-Gaussian multi-modal distributions. This paper describes two new algorithms that integrate the marginalized particle filter and maximum a-posterior particle filter and demonstrates challenging cases where the proposed algorithms outperform the conventional marginalized particle filter using both synthetic and real electrocardiogram signals.
展开▼
