The QRS complexes and ventricular beats in an electrocardiogram represent the depolarization phenomenon of the ventricles and yield useful information about their behavior. Beat detection is a procedure preceding any kind of ECG processing and analysis. For morphological analysis this is the reference for detection of other ECG waves and parameter measurements. Rhythm analysis requires classification of QRS and other ventricular beat complexes as normal and abnormal. Real-time ventricular beat detection is essential for monitoring of patients in critical heart condition.

Correct beats recognition is impeded by power-line interference, electromyogram noise and baseline wander often present in the ECG signal.

In long-term monitoring electrode impedance can increase considerably, resulting in very low signal-to-noise ratio, which can make detection practically impossible in a single lead. Therefore, usually two or three leads are used for monitoring 1.

Friesen et al. 2 have presented a comparison of nine QRS detection algorithms, based on: i) amplitude and first derivative, ii) first derivative only, iii) first and second derivative, and iv) digital filtering. Daskalov et al. 3 applied these algorithms to selected signals containing records with pronounced baseline drift. The results were unsatisfactory, which was probably due to the use of fixed detection thresholds, whereas adaptive ones would be more appropriate.

Poli et al. 4 used a generic algorithm for QRS detection. The complexes were emphasized with respect to the rest of the signal by polynomial filters and compared to an adaptive threshold. The authors reported 99.60 % sensitivity (Se) and 99.51 % specificity (Sp) with the MIT-BIH Arrhythmia Database. The method is inapplicable in real-time.

Afonso et al. 5 proposed hardware filter banks for ECG signal decomposition, where several parameters were independently computed and combined in a decision rule. The authors reported Se = 99.59 % and Sp = 99.56 % for their real-time, single-channel beat detection algorithm tested with the MIT-BIH Arrhythmia Database.

Dotsinsky and Stoyanov 6 developed a heuristic, pseudo-real-time algorithm for ventricular beat detection for single-channel ECG, based on steep edges and sharp peaks evaluation criteria. They reported Se = 99.04% and Sp = 99.62%, obtained with two channel recordings from AHA and MIT-BIH Arrhythmia Database

Moraes et al. 1 combined logically two different algorithms working in parallel – the first has been taken from the work of Englese and Zeelenberg 7 and the other was based on Pan and Tompkins 8, and Ligtenberg and Kunt 9. Moraes et al. 1 reported Se = 99.22 % and Sp = 99.73 % after having excluded records of patients with pacemaker. After excluding a few more recordings 108, 200, 201 and 203, containing high amplitude noise (according to the authors), the statistical indices rises to Se = 99.56 % and Sp = 99.82 %.

Li et al. 10 have used wavelet transforms for detection. They reported 0.15 % false detections out of 46 files from the MIT-BIH Arrhythmia Database, but with exclusion of files 214 and 215. In addition, we found some errors in their Table II. After correction, the reported accuracy slightly decreased.

The large variety of QRS detection algorithms, and the continuous efforts for their enhancement, proves that universally acceptable solution has not been found yet. Difficulties arise mainly from the huge diversity of the QRS complex waveforms and the noise and artifacts accompanying the ECG signals.

All 48 ECG recordings of MIT-BIH Arrhythmia database were used, without exception. Each one has a duration of 30 min and includes two leads – the modified limb lead II and one of the modified leads V1, V2, V4 or V5 11. The sampling frequency is 360 Hz with resolution 5 μV/bit. Two cardiologists have annotated all beats. Approximately 70 % of the beats are annotated as Normal. Four of the records are from patients with pacemakers.

The American Heart Association (AHA) database was also considered, during the evaluation of the method, mostly due to the fact that it contains patients with premature ventricular beats of contraction of R-over-T type, some of them very difficult to detect because of their closeness to the previous complex. Statistical indices for this database are not derived, because they can be compared with limited number of articles working with AHA. The database consists of 80 recordings: 2-leads, 250 Hz sampling rate and 5 μV/bit resolution.

The differentiated and summed signals from L leads are compared to the absolute value of a threshold MFR = M + F + R – a combination of three independent adaptive thresholds, where:

• M – Steep-slope threshold;

• F – Integrating threshold for high-frequency signal components;

• R – Beat expectation threshold.

Two algorithms were developed:

Algorithm 1 detects at the current beat.

Algorithm 2 Pseudo-real-time detection with additional triggering of potentially missed heart beat in the last interval by RR interval analyses.

The algorithms are self-adjusting to the thresholds and weighting constants, regardless of resolution and sampling frequency used. They operate with any number L of ECG leads, self-synchronize to QRS or beat slopes and adapt to beat-to-beat intervals.

Normally the statistical indices Se and Sp are derived from the following parameters: correctly detected beats TP (true positive), falsely detected beats FP (false positive) and undetected beats FN (false negative). In addition, we used two parameters, adopted from Dotsinsky and Stoyanov 6, as described below.

SP – shifted positive error was introduced in order to explain cases like the one shown in Fig. 6. Here the algorithm made a false positive error before the 3rd QRS and missed the following QRS. Formally, this is a false positive error, immediately followed by a false negative. However, if the time interval between these two detections is reasonably short, for example ≤ 200 ms, we accepted this as one error only, labeled as Shifted False Positive Error (SP).

Another example of SP error as a result of artifacts just before the normal complexes is shown in Fig. 7.

SN – shifted negative error was assumed by the same principle as SP, but in the opposite sense. It also included twin FN+FP error occurring within 200 ms. The first incoming FP or FN error of the shifted is defining it as SP or SN.

The logic of using shifted errors (instead of FP and FN or FN and FP in cases when they appear within 200 ms of each other) is that thus the total number of beats in a record retains its value. Otherwise it would change depending on the type and number of errors and thus impede correct computation of Se and Sp.

The sensitivity Se is calculated by summing FN SN, while the specificity Sp – by summing FP+SP.

The method was developed in Matlab. All 48 recordings from the MIT-BIH Arrhythmia database, without any exception, were used for testing the two algorithms.

The processed files containing detection marks were automatically compared with the original MIT-BIH annotated beats by specially designed software. It shows all cases where the annotation and detection marks differ by more than 60 ms. These cases were examined by an independent expert, thus excluding possible author's influence.

The results for the two algorithms are presented in Table 1.

The proposed algorithms for real-time and pseudo-real-time implementation are adaptive, independent of thresholds and constants values. They are self-synchronized to the QRS steep slope and the heart rhythm, regardless of the resolution and sampling frequency used. Due to the integration threshold, the algorithms are practically insensitive to electromyogram and similar high-frequency noise.

The algorithms can operate with one, two or more leads, using a combined lead signal derived from the sum of absolute values of the differentiated lead signals.

The statistical indices are higher than, or comparable to those, cited in the scientific literature.