Ivaylo Christov

The aim of the study is to assess the performance of an automated method for Q-onset and T-end delineation, as well as QT measurement with the use of a 'gold standard' of a manually created reference database. The data to be used comprise 548 recordings of the PTB Diagnostic ECG Database. The ECG signals are preprocessed, suppressing power-line interference, electromyographic noise and baseline drift according to our previously published investigations, guaranteeing accurate Q-onset and T-end locality preservation. Our method for automatic detection of Q-onset and T-end is based on the minimum value of the angle between two segments having a common mid point and equal lengths of 10 ms. The minimum of the angle is searched in defined time intervals delimited separately for the Q-onset and T-end. All measurements are performed on lead II only. Mean +/- standard deviations are obtained for 95% of the recordings: -0.08 +/- 2.71 for Q-onset, 5.10 +/- 9.22 for T-end and 4.40 +/- 9.93 for QT interval, as well as for 100% of the recordings: 0.46 +/- 4.84 for Q-onset, 1.28 +/- 16.75 for T-end and 0.83 +/- 16.67 for QT interval.

The QT interval and the QT dispersion are currently a subject of considerable interest. Cardiac repolarization delay is known to favor the development of arrhythmias. The QT dispersion, defined as the difference between the longest and the shortest QT intervals or as the standard deviation of the QT duration in the 12-lead ECG is assumed to be reliable predictor of cardiovascular mortality. The seventh annual PhysioNet/Computers in Cardiology Challenge, 2006 addresses a question of high clinical interest: Can the QT interval be measured by fully automated methods with accuracy acceptable for clinical evaluations? The PTB Diagnostic ECG Database was given to 4 cardiologists and 1 biomedical engineer for manual marking of QRS onsets and T-wave ends in 458 recordings. Each recording consisted of one selected beat in lead II, chosen visually to have minimum baseline shift, noise, and artifact.In cases where no T wave could be observed or its amplitude was very small, the referees were instructed to mark a 'group-T-wave end' taking into consideration leads with better manifested T wave.A modified Delphi approach was used, which included up to three rounds of measurements to obtain results closer to the median. A total amount of 2*5*548 Q-onsets and T-wave ends were manually marked during round 1. To obtain closer to the median results, 8.58 % of Q-onsets and 3.21 % of the T-wave ends had to be reviewed during round 2, and 1.50 % Q-onsets and 1.17 % T-wave ends in round 3. The mean and standard deviation of the differences between the values of the referees and the median after round 3 were 2.43 +/- 0.96 ms for the Q-onset, and 7.43 +/- 3.44 ms for the T-wave end. A fully accessible, on the Internet, dataset of manually measured Q-onsets and T-wave ends was created and presented in additional file: 1 (Table 4) with this article. Thus, an available standard can be used for the development of automated methods for the detection of Q-onsets, T-wave ends and for QT interval measurements.

The prompt and adequate detection of abnormal cardiac conditions by computer-assisted long-term monitoring systems depends greatly on the reliability of the implemented ECG automatic analysis technique, which has to discriminate between different types of heartbeats. In this paper, we present a comparative study of the heartbeat classification abilities of two techniques for extraction of characteristic heartbeat features from the ECG: (i) QRS pattern recognition method for computation of a large collection of morphological QRS descriptors; (ii) Matching Pursuits algorithm for calculation of expansion coefficients, which represent the time-frequency correlation of the heartbeats with extracted learning basic waveforms. The Kth nearest neighbour classification rule has been applied for assessment of the performances of the two ECG feature sets with the MIT-BIH arrhythmia database for QRS classification in five heartbeat types (normal beats, left and right bundle branch blocks, premature ventricular contractions and paced beats), as well as with five learning datasets-one general learning set (GLS, containing 424 heartbeats) and four local sets (GLS+about 0.5, 3, 6, 12 min from the beginning of the ECG recording). The achieved accuracies by the two methods are sufficiently high and do not show significant differences. Although the GLS was selected to comprise almost all types of appearing heartbeat waveforms in each file, the guaranteed accuracy (sensitivity between 90.7% and 99%, specificity between 95.5% and 99.9%) was reasonably improved when including patient-specific local learning set (sensitivity between 94.8% and 99.9%, specificity between 98.6% and 99.9%), with optimal size found to be about 3 min. The repeating waveforms, like normal beats, blocks, paced beats are better classified by the Matching Pursuits time-frequency descriptors, while the wide variety of bizarre premature ventricular contractions are better recognized by the morphological descriptors.

Background: Modern biomedical amplifiers have a very high common mode rejection ratio. Nevertheless, recordings are often contaminated by residual power-line interference. Traditional analogue and digital filters are known to suppress ECG components near to the power-line frequency. Different types of digital notch filters are widely used despite their inherent contradiction: tolerable signal distortion needs a narrow frequency band, which leads to ineffective filtering in cases of larger frequency deviation of the interference. Adaptive filtering introduces unacceptable transient response time, especially after steep and large QRS complexes. Other available techniques such as Fourier transform do not work in real time. The subtraction procedure is found to cope better with this problem. Method: The subtraction procedure was developed some two decades ago, and almost totally eliminates power-line interference from the ECG signal. This procedure does not affect the signal frequency components around the interfering frequency. Digital filtering is applied on linear segments of the signal to remove the interference components. These interference components are stored and further subtracted from the signal wherever non-linear segments are encountered. Results: Modifications of the subtraction procedure have been used in thousands of ECG instruments and computer-aided systems. Other work has extended this procedure to almost all possible cases of sampling rate and interference frequency variation. Improved structure of the on-line procedure has worked successfully regardless of the multiplicity between the sampling rate and the interference frequency. Such flexibility is due to the use of specific filter modules. Conclusion: The subtraction procedure has largely proved advantageous over other methods for power-line interference cancellation in ECG signals.

Background A specific problem using the public access defibrillators (PADs) arises at the railway stations. Some countries as Germany, Austria, Switzerland, Norway and Sweden are using AC railroad net power-supply system with rated 16.7 Hz frequency modulated from 15.69 Hz to 17.36 Hz. The power supply frequency contaminates the electrocardiogram (ECG). It is difficult to be suppressed or eliminated due to the fact that it considerably overlaps the frequency spectra of the ECG. The interference impedes the automated decision of the PADs whether a patient should be (or should not be) shocked. The aim of this study is the suppression of the 16.7 Hz interference generated by the power supply of the railway systems. Methods Software solution using adaptive filtering method was proposed for 16.7 Hz interference suppression. The optimal performance of the filter is achieved, embedding a reference channel in the PADs to record the interference. The method was tested with ECGs from AHA database. Results The method was tested with patients of normal sinus rhythms, symptoms of tachycardia and ventricular fibrillation. Simulated interference with frequency modulation from 15.69 Hz to 17.36 Hz changing at a rate of 2% per second was added to the ECGs, and then processed by the suggested adaptive filtering. The method totally suppresses the noise with no visible distortions of the original signals. Conclusion The proposed adaptive filter for noise suppression generated by the power supply of the railway systems has a simple structure requiring a low level of computational resources, but a good reference signal as well.

An analysis of electrocardiographic pattern recognition parameters for premature ventricular contraction (PVC) and normal (N) beat classification is presented. Twenty-six parameters were defined: 11 x 2 for the two electrocardiogram (ECG) leads, width of the complex and three parameters derived from a single-plane vectorcardiogram (VCG). Some of the parameters include amplitudes of maximal positive and maximal negative peaks, area of absolute values, area of positive values, area of negative values, number of samples with 70% higher amplitude than that of the highest peak, amplitude and angle of the QRS vector in a VCG plane. They were measured for all heartbeats annotated as N or PVC in all 48 ECG recordings of the MIT-BIH arrhythmia database. Two reference sets for the Kth nearest-neighbours rule were used-global and local. The classification indices obtained with the global reference set were 75.4% specificity and 80.9% sensitivity. Using the local reference set we increased the specificity to 96.7% and the sensitivity to 96.9%. The achieved specificity and sensitivity are comparable with, and greater than, the results reported in the literature.

Background: QRS and ventricular beat detection is a basic procedure for electrocardiogram (ECG) processing and analysis. Large variety of methods have been proposed and used, featuring high percentages of correct detection. Nevertheless, the problem remains open especially with respect to higher detection accuracy in noisy ECGs Methods: A real-time detection method is proposed, based on comparison between absolute values of summed differentiated electrocardiograms of one of more ECG leads and adaptive threshold. The threshold combines three parameters: an adaptive slew-rate value, a second value which rises when high-frequency noise occurs, and a third one intended to avoid missing of low amplitude beats. Two algorithms were developed: Algorithm 1 detects at the current beat and Algorithm 2 has an RR interval analysis component in addition. 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. Results: The algorithms were tested by an independent expert, thus excluding possible author's influence, using all 48 full-length ECG records of the MIT-BIH arrhythmia database. The results were: sensitivity Se = 99.69 % and specificity Sp = 99.65 % for Algorithm 1 and Se = 99.74 % and Sp = 99.65 % for Algorithm 2. Conclusion: The statistical indices are higher than, or comparable to those, cited in the scientific literature.

Detection and classification of ventricular complexes from a limited number of ECG leads is of considerable importance in critical care or operating room patient monitoring. Beat-to-beat detection allows the heart rhythm evolution to be followed and various types of arrhythmia to be recognized. A quantitative analysis is proposed of pattern recognition parameters for classification of normal QRS complexes and premature ventricular contractions (PVC). Twenty-six parameters have been defined: the width of the QRS complex, three vectorcardiogram parameters and 11 from two ECG leads. These parameters include: amplitudes of positive and negative peaks, area of positive and negative waves, various time-interval durations, amplitude and angle of the QRS vector, etc. They are measured for all QRS complexes annotated as 'normals' and 'PVCs' from the 48 ECG recordings of the MIT-BIH arrhythmia database. Neural networks (NN) are shown to be a useful instrument for the analysis of large quantities of parameters. Separate ranking of any parameter and homogeneous group ranking (amplitude, area, interval, slope and vector) were performed. From the two ECG leads, the first three ranked parameter groups for clustering of PVCs are amplitude, slope and interval, while for N clustering they are vector, amplitude and area. Considering the entire parameter set, we obtained N = 99.7% correct detection of normal QRS complexes and PVC = 98.5% of premature ventricular complexes. The study also shows that simultaneous analysis of two ECG channels yields better accuracy compared to using a single channel: the improvement is 0.1% in the classification of N beats and 4.5% for PVC beats.