Bioelectrical signals are often pulse-shaped with superimposed interference signals. In this context, accurate identification of features such as pulse onsets, peaks, ampli- tudes, and duration is a frequent problem. In this paper, we present a versatile method of rather low computational complexity to robustly identify such features in real-world signals. For that, we take use of two straight-line models fit to the observations by minimizing a quadratic cost term, and then identify desired features by tweaked likelihood measures. To demonstrate the idea and facilitate access to the method, we provide examples from the field of cardiology.

Signal-to-peak-interference ratio (SPIR) optimal filters are template matching filters with peak interference suppression properties. Such max-SPIR filters are used in multi-pattern recognition problems, such as neural spike sorting in micro-electrode array probes, where cellular action potentials need to be detected and clustered according to their firing neuron cells. In high-density probes with hundreds of channels, such max-SPIR filter banks can require unacceptable high computational resources, in particular for applications with real-time demands and/or on-probe spike sorting. In this paper, we present a computationally attractive substitute for max-SPIR filters by recursively computed Autonomous Linear State Space Model (ALSSM) filters. In our approach, we approximate the impulse response of max-SPIR filters by low order ALSSMs and perform the signal convolution in the new, low-dimensional ALSSM vector space. We demonstrate our method on real neural recordings from high-density probes and show only minimal loss in detection quality while the computational complexity drops by up to a factor 10.