By fitting a nonlinear model for generating realistic ECG to waveforms collected from a healthy human subject (described at CinC 2004), and using a nonlinear least-squares optimization procedure, the authors demonstrate that significant points (P, Q, R, S, and T) on the ECG can be determined to an arbitrary accuracy. The model-fitting routine runs in real-time on a 3GHz PC. Coloured 1/f-like noise (where the frequency f has been raised to a power, beta)is then added to the ECG in order to evaluate the fitting accuracy under a variety of recording conditions. A method for determining noise levels (and colour) in real ECGs using the residual of a singular valued decomposition is described. Furthermore, a method for evaluating the filter is described which allows an application-specific evaluation of the filter in terms of the distortion in the QRS width and amplitude, the ST-level, the QT interval, the PR-interval, and the fiducial point location. Using these methods, we found that the model-based filter introduced clinically insignificant distortion in QT interval and QRS width (<40ms and 20ms respectively) for signal/noise ratios (SNRs) as low as 0 dB if beta < 2, with fiducial point jitter < 1 ms for SNR <= 2 dB, and stable ST level measurements for SNR > 12 dB. PR interval is shown to be more sensitive to noise due to the low amplitude nature of the P-wave. In general, the filter performance improves as beta decreases (i.e., as the noise whitens).