The Physionet QT Database: Study on the Reliability of P-wave Manual Annotations under Noisy Recordings

Abstract

Thanks to its manual annotations, the PhysioNet QT database (QTDB) has been widely used as the reference of ECG delineators. However, a significant percentage of its annotations have been reported as inaccurate. Thus, any precise ECG delineator will never be able to meet, without error, all its annotations. The present work analyzes these inaccuracies and also how noise altered the final timing of annotations. As this effect is higher for low amplitude waveforms, P-waves were studied through a robust P-wave delineator. Its delineation results were compared with manual annotations under two scenarios. Firstly, a direct comparison without ECG denoising was performed. Secondly, the P-waves were delineated after efficient Wavelet-based denoising. Results showed that automatic annotations were closer to manual annotations for noisy ECGs and farther in the case of denoised ECGs, thus proving that noise altered the timing of manual annotations. An unreal improvement in delineation performance for noisy ECGs was obtained for P-wave onset, peak and offset in 45.83%, 57.29% and 56.25% of the recordings, respectively. Thus, to improve delineators reliability, either the need to review the QTDB annotations or its replacement by a better annotated database are suggested.

ECG Beat Representation and Delineation by Means of Variable Projection

“…Consequently, we evaluated the performance of our method using the Physionet QTDB [16], [17] to demonstrate that it delivers results that are comparable to those of the latest ECG beat delineation algorithms. Despite several limitations [42], the QTDB remains one of the best options for testing the robustness of a delineation algorithm; furthermore, it has been used by several research groups to compare delineation results. In the next three paragraphs, we thus describe the delineation method, the QTDB, and our results.…”

“…In the case of P on , for instance, this would mean that the bias must be smaller than 25 ms, while the σ e should be smaller than 30 ms for a recording to be assigned to group I. For group II the difference between expert and algorithm annotations is greater, which does not automatically mean poor performance, because in some cases even experts disagree on the true value of the characteristic point [42]. Groups III and IV have a high standard deviation, which indicates varying algorithm (or expert) annotations for similar beats within a recording and is, of course, not desired at all.…”

“…First, it contains expert annotations for more than 3000 beats of 105 2-channel ECG recordings, which provides a large variety of beat morphologies and allows a method's robustness against pathological / atypical wave forms to be estimated. Second, it has been used extensively by other research groups, and therefore we can easily estimate the general delineation quality by comparing its results to those provided in the literature [42]. Lastly, our previous approach, where we used adaptive Hermite functions without spline interpolation and sigmoidal functions for ECG wave delineation [9], was also evaluated on this database.…”

Thorough Assessment of a P-wave Delineation Algorithm Through the Use of Diverse Electrocardiographic Databases

“…The performance of these methods has been mostly validated making use of the QT database, which is freely available at PhysioNet [7] and also contains manual annotations from experts on the beginning, the peak and the end of the Pwaves [8]. However, in the last years several authors have evidenced some major limitations of this database, such as the lack of information on the specific channel used for delineating each P-wave, the diversity of criteria used in the annotation process, or the noise effect on the result of the manual annotation [9], [10]. As a consequence, they have also suggested the need for a more robust validation of the P-wave delineation algorithms on new carefully annotated databases [11].…”