Non linear filtering for the estimation of the respiratory component in heart rate

Three quite general nonlinear adaptive filters were tested both for cancelling/estimating the contribution of the respiratory activity in the RR time series and for estimating a nonlinear model of RR time series. The Volterra 2nd order filter, bilinear system, tree-structured piecewise linear filter were considered. The algorithms used for the convergence are similar to known Recursive Least Squares and to Least Mean Square. Absolute error, square error or fourth power error were used in the "optimum" criterion. The rest of the nonlinear adaptive filters on time series of patients under different pathophysiological conditions (normal, hypnotic relaxation, periodic breathing, transplant) showed no significant improvement in comparison with the linear filters. Although nonlinear filters performed better in specific intervals than linear filters, they were found to be more sensitive to noise and artifacts.<>