The use of the differential steepest descent algorithm for adaptive template matching

The differential steepest descent algorithm is presented in a form useful for template matching of biomedical signals. A template pattern is adaptively weighted toward optimally matching an input pattern in the least squares sense. Parameters such as gain, DC bias, phase, and sampling interval weights are adjusted iteratively, according to the sum of squares error obtained by subtraction of template from input pattern, point by point. For biomedical pattern recognition, the template pattern may be obtained either from experimental data or from model equations. The technique is relevant to several types of real-time biomedical applications: (1) tracking of pattern parameters over time, (2) preprocessing, such as obtaining the best window and/or normalization of an input pattern before implementation of optimal features selection procedures, and (3) the least squares error at convergence to the optimal weight vector is itself useful information for pattern recognition. The technique is used to match a blood pressure pulse taken from dog data with three harmonics of a model blood pressure wave. Stability and convergence properties of the technique are shown, and suggestions are made for matching patterns that have undergone nonlinear transformations of shape