This article deals with the problem of on-line wave detection in non-stationary signals, through a polynomial approximation based on genetics algorithms. This approximation is estimated by minimising a nonlinear error function, described by the circle equation. We apply this approach to detect sleep spindle waveforms in electroencephalograms (EEG) and R wave and ectopics beats in electrocardiograms (ECG). The proposed method makes it possible to recognise time-variant waveform with a simple decisional parameters.