Development and Evaluation of a QT Interval Algorithm Using Different ECG Databases

Dieter Hayn^a, Günter Schreier^a, Suave Lobodzinski^b

^aARC Seibersdorf research GmbH, Biosignal Processing and Telemonitoring, Graz, Austria
^bCalifornia State University, Long Beach, CA, USA

Correspondence: D Hayn, ARC Seibersdorf research GmbH, Biosignal Processing and Telemonitoring,
Grottenhofstrasse 3, 8053 Graz, Austria. E-mail: dieter.hayn@arcs.ac.at, phone +43 316 265158, fax +43 316 265155

1.        Introduction

Observing the QT interval from the ECG is a well established method to check for a possible arrhythmogeneity of newly developed therapeutic agents. It was the aim of this study to develop an algorithm for automatically calculating the onset of the QRS complex and the offset of the T wave using different kinds of ECG signals obtained from two databases and to find out, whether methods developed with one kind of signal can directly be applied to signals of a somehow different type.

2.        Material and Methods

For pre-processing of the raw signals we used our existing software package that was written in Matlab 6.5 (The MathWorks, Inc., Natick, MA 1760-2098, USA) and contains software for QRS detection and correctly identifying single beats [Schreier et al., 2001]. Coarse wave detection for each channel was based on the definition of a short time window that was pulled over a single beat. Within this window we calculated the ranges of the signal amplitudes and compared these ranges with certain threshold values as illustrated in Fig. 1. The time windows, within which we tried to find the wave onset and offset, were obtained from the signals of the QT database on PhysioNet, which is described in [Laguna et al., 1997].

After this coarse wave detection three independent methods for finding the exact beginnings and endings of the characteristic waves were used and their results were averaged. Finally we combined the results obtained from all channels by calculating a confidential parameter for each point of each channel. The final point depended on the points found for each channel and their confidential parameters.

The algorithm was developed and optimized using the QT database and then applied to the CSE multilead database prospectively and without any further tuning. We only adapted it to the different file and signal formats. Both databases contain expert annotations for QRS onset and T offset. However, they differ with respect to a number of factors, among them the number of channels (QT: 2, CSE: 15) and the number of expert annotations per point (QT: 1 or 2, CSE: 5).

We compared the time points found by our algorithm with those annotated by the experts in both databases. Where there was more than one expert annotation per beat we took the mean value of all experts’ annotation as a reference. Since in the CSE database only the first beat of each signal is annotated, validation of the PhysioNet database was also done using the first annotated beat in each signal only.

Figure 1. Coarse T offset detection. The range curve undershoots the threshold value twice. The first undershoot was correctly interpreted as a maximum value of the original signal and the second undershoot was chosen as the coarse estimation for T offset.

3.        Results

Our algorithm was able to detect 100% of the analyzed QRS onset and T offset points of both databases. Mean value and standard deviation of the distances in between the time points found by our algorithm and the mean values of expert’s annotations are summarized in Table 1.

Table 1.    Deviation in between characteristic points found by our algorithm and those annotated by experts for the QT database on PhysioNet and the CSE multilead database, respectively. Column four shows the mean range of time points annotaded by different experts for each point of both databases.

4.        Discussion

Standard deviation values within the mean expert ranges indicate the stability of the developed QT algorithm. The significant bias found for the calculations of the T offset values in the CSE database, however, indicates a systematical error in wave detection with respect to the CSE database. Manual inspection indicated that the availability of more channels indeed leads to later T offset points since experts tend to set their markers at the latest point found in all channels. Similar observations for other algorithms have also been found by the CSE study [Willems et al., 1985].

We conclude that using ECG databases with different characteristics may reveal otherwise unnoticed issues and finally lead to the development of a more generally applicable algorithm.

References

Laguna P, Mark RG, Goldberger A, Moody GB. A database for evaluation of algorithms for measurement of QT and other waveform intervals in the ECG. Computers in Cardiology, 24: 673-676, 1997.

Schreier G, Kastner P, Marko W. An automatic ECG processing algorithm to identify patients prone to paroxysmal atrial fibrillation. Computers in Cardiology, 28: 133-135, 2001.

Willems JL, Arnaud P, Van Bemmel JH. Assessment of the performance of electrocardiographic computer programs with the use of a reference database. Circulation, 71(3): 523-534, 1985.