Model-based determination of QT intervals

A method is presented to determine the QT interval by fitting a nonlinear artificial ECG model to segmented regions of a human ECG. The model consists of a set of temporally Gaussian functions with different widths and heights. These parameters are fitted to a given ECG (segmented around the QRS complex to include the P and T wave) using a nonlinear least squares optimization routine. The Q onset and T offset can be determined precisely (in a statistical sense) from the parameters of the Gaussian. Since the fitted waveform contains no noise, the differential is smooth. Waveform boundaries can also be determined by searching for the minimum of a differential. Furthermore, the residual error provides an estimate of the confidence in the fit, and hence, the derived QT interval. Using the human expert-annotated PhysioNet QT database, various QT interval estimation schemes were compared using the model-fitted ECG to find an optimal marker of the QT interval. It was found that humans are inconsistent and almost always under-estimate the T-offset (if defined to be the end of any repolarization). This is probably due to the truncation of any human estimation when the T wave tail is consumed by noise. We therefore propose an alternative QT end point. Finally, an entry based on the most favourable technique was submitted in the PhysioNet / Computers in Cardiology Challenge 2006; QT Interval Measurement, which is intended to produce a comparison of several automatic and human annotators on the Physikalisch-Technische Bundesanstalt diagnostic ECG database. A follow-up paper to address differences between those generated by our method and the consensus of the other entries will be submitted shortly.