Development of a toolbox for electrocardiogram-based interpretation of atrial fibrillation

Introduction 

Atrial fibrillation (AF) is the most common type of human cardiac arrhythmia. Its clinical importance is due to increasing risk of embolic complications, stroke, and heart failure. It is an important clinical entity because of the increasing risk of morbidity and mortality.1, 2 Atrial fibrillation health costs amount to around 442 million US dollars (around 961 million US dollars with nursing home care and secondary admissions) for a total of 601 000 individuals with AF in the UK. Statistics reveal some 2 500 000 strokes annually worldwide (United States: 500 000). In approximately 30% of these cases, AF is found. Its prevalence is expected to double in the next 50 years.

The diagnosis of AF as such has been based mainly on visual inspection of the surface electrocardiogram (ECG).3 Due to the much higher amplitude of the electrical ventricular activity (VA) on the surface ECG, suppression of the ventricular involvement is crucial in the noninvasive study of AF. This suppression has a direct influence on the quality of the subsequent signal processing leading to the interpretation of the view on the resulting ECG activity. A proper VA cancellation can improve the visual distinction between rhythms such as AF, atrial flutter, and different atrioventricular blocks by displaying separately the atrial and ventricular signal components, as well as in software-based decision procedures.

Two approaches are generally used to perform VA suppression: source separation algorithms and matched template subtraction. Source separation algorithms try to find uncorrelated components using principal component analysis, or to find independent components in an instantaneous linear mixture using independent component analysis. The principal component analysis has previously been used to monitor the effects of drugs4 and assess the effects of linear left atrial ablation.5 The independent component analysis has been applied to obtain ECG signals devoid of VA involvement.6 In these algorithms, the components obtained are not directly interpretable in terms of the electrophysiology involved and its expression in the signals generated at specific thorax locations (electrode positions). The other well-known approach is based on the subtraction of matched templates of the QRST complexes. This approach used the fact that in an individual patient, ventricular complexes generally exhibit a limited number of distinct waveforms only. Averages of such distinct complexes are used for building up templates, which are subsequently subtracted from the signals of the corresponding beats. This standard average beat subtraction method is labeled ABS. An adapted ABS method, published by Lemay et al7 in 2007, is referred to here as the ABS-EPFL procedure. In ABS methods, several beats are needed for each QRST morphology to create the templates. In clinical practice, where the standard ECG consists of 10-second recordings, high-quality templates may be difficult to create. A novel method for suppressing the ventricular signal, permitting shorter records, is the single beat algorithm, referred to here as SB-EPFL.7

In the work reported on in this article, we used the methods published by Lemay et al7 in 2007 and implemented the subtraction algorithm into an industrially applicable software package called the AF-Toolbox. The transition of the algorithmic part from a research tool, developed in Matlab (Mathworks, Natick, MA), to an implementation in American National Standards Institute (ANSI) C was the most demanding task. The results obtained by using the dynamic link library were compared with those based on the published research algorithms,7 aimed at testing for any loss in performance. In addition, the article reports on the implementation of some features, derived from these “cleaned up” signals, such as atrial and ventricular heart rate and the ratio between these 2 quantities.

Methods and materials 

Algorithm transition/adaptation 

The ABS-EPFL and SB-EPFL algorithms for QRST suppression7 needed to be adapted to be used in a clinical ECG recorder. The change in precision from double (64-bit floating point numbers) to integer precision had to be taken into account. Also, an ECG machine is not comparable with a standard personal computer from the point of view of available processing speed and memory. In addition, conversion of the algorithms from a Matlab script to an industrial applicable ANSI C code may easily give rise to performances loss. The latter obviously needed to be minimized.

The resulting AF-Toolbox algorithm is characterized by an optimized sequence of processing steps. The preprocessing comprises a baseline correction (using cubic splines).

The next unit reduces the F waves that are present in the case of atrial flutter. It aims at preventing any unintended subtraction of the F waves (if present) during the suppression of the QRST complexes.8 This step is essential because flutter waves represent atrial activity (AA) that is usually highly correlated with VA.

The main processing comprises units for either average beat or dominant T wave subtraction9 applied to the individual 8 independent leads: II, III, V1,…V6. In the ABS-EPFL procedure, up to 4 different beat classes/templates are used. In the SB-EPFL procedure, the signal of an individual beat after the QRST complex is fitted by a linear combination of 3 derivatives of the dominant T wave (the derivatives of orders 0, 1, and 2). The fitting is performed for each beat, applied to the intervals from the end of the QRS complexes until the onset of the next QRS complexes. The signals during QRS intervals are filled in by means of a Fourier interpolation of the “cleaned up” segments before and after these QRS intervals. The resulting features comprise atrial and ventricular rates, their ratio (Fig. 1), a suggested classification of the arrhythmia involved, and, in the case of atrial flutter, its characterization (direction of revolution: clockwise or anticlockwise).

View Large Image Download to PowerPoint

Fig. 1. Picture of the AF-Toolbox. The calculating part of the AF-Toolbox is programmed in C (DLL, or dynamic link library); for research and development phase, Matlab was used as GUI (graphic user interface) and will be replaced by the ECG machine's corresponding GUI. This is a typical AF case.

Results 

The separation algorithm was implemented in ANSI C (as dynamic link library) and called by a graphic user interface (Matlab-GUI). This is used during the development phase only; it will be replaced by an ECG device corresponding GUI. An example of a typical AF case as viewed from the interface is shown in Fig. 1.

On the synthetic signals, the performance of the AF-Toolbox algorithm can be quantified as follows: NMSE of 0.75 ± 0.11 (mean ± SD) during the QRS intervals. The standard ABS yielded a mean NMSE of 2.99 ± 2.60, the ABS-EPFL of 1.31 ± 0.32, and the SB-EPFL NMSE of 1.16 ± 0.12 (Table 1). Outside the QRS complexes (JQ intervals), the AF-Toolbox algorithm yielded a mean NMSE of 0.27 ± 0.08, the standard ABS of 0.32 ± 0.22, the ABS-EPFL of 0.18 ± 0.08, and the SB-EPFL of 0.65 ± 0.29 (Table 2). The percentage of significant QRS residuals detected on the 180 clinical 5-minute ECG signals is 5.7% for the AF-Toolbox, 53.8% for the standard ABS, 12.0% for the ABS-EPFL, and 3.9% for the SB-EPFL (Table 3). A typical example of a 10-second ECG signal from the simulated test set is shown in Fig. 2. In the latter, the AA signal was magnified to provide a better view of the differences between the original and the estimated AA.

Table 1. NMSE of the QRS intervals for each lead of the 18 simulated ECGs

	NMSE	ABS	ABS-EPFL	SB-EPFL	AF-Toolbox
	VR	1.58 ± 1.45	1.31 ± 0.19	1.12 ± 0.09	0.76 ± 0.11
	VL	1.90 ± 1.19	1.43 ± 0.22	1.28 ± 0.17	0.81 ± 0.10
	V1	1.06 ± 0.36	1.09 ± 0.19	1.13 ± 0.12	0.71 ± 0.09
	V2	1.96 ± 0.82	1.17 ± 0.22	1.17 ± 0.09	0.74 ± 0.11
	V3	4.16 ± 3.58	1.46 ± 0.54	1.14 ± 0.11	0.75 ± 0.11
	V4	6.08 ± 7.79	1.41 ± 0.60	1.16 ± 0.17	0.75 ± 0.11
	V5	4.33 ± 3.90	1.30 ± 0.36	1.18 ± 0.14	0.75 ± 0.12
	V6	2.82 ± 1.67	1.32 ± 0.23	1.13 ± 0.12	0.76 ± 0.11
	Average	2.99 ± 2.60	1.31 ± 0.32	1.16 ± 0.12	0.75 ± 0.11

NMSE indicates normalized mean square error.

Table 2. NMSE of the JQ intervals for each lead of the 18 simulated ECGs

	MSE	ABS	ABS-EPFL	SB-EPFL	AF-Toolbox
	VR	0.30 ± 0.21	0.16 ± 0.04	0.60 ± 0.28	0.24 ± 0.07
	VL	0.35 ± 0.21	0.27 ± 0.12	1.08 ± 0.91	0.34 ± 0.08
	V1	0.27 ± 0.23	0.12 ± 0.04	0.45 ± 0.10	0.23 ± 0.07
	V2	0.31 ± 0.22	0.16 ± 0.04	0.80 ± 0.22	0.26 ± 0.09
	V3	0.35 ± 0.24	0.22 ± 0.13	0.62 ± 0.21	0.26 ± 0.09
	V4	0.32 ± 0.24	0.20 ± 0.15	0.61 ± 0.35	0.26 ± 0.10
	V5	0.31 ± 0.21	0.16 ± 0.07	0.48 ± 0.15	0.25 ± 0.08
	V6	0.35 ± 0.21	0.18 ± 0.06	0.55 ± 0.11	0.28 ± 0.08
	Average	0.32 ± 0.22	0.18 ± 0.08	0.65 ± 0.29	0.27 ± 0.08

MSE indicates mean square error.

Table 3. Percentage of significant QRS residuals detected on 180 clinical 5-minute ECG results

	% of QRS	ABS	ABS-EPFL	SB-EPFL	AF-Toolbox
	VR	33.5%	5.8%	3.4%	11.9%
	VL	44.6%	6.8%	5.5%	9.8%
	V1	34.0%	4.6%	3.1%	3.9%
	V2	59.4%	12.2%	4.8%	4.5%
	V3	59.0%	12.4%	2.7%	3.7%
	V4	58.7%	14.2%	3.7%	4.9%
	V5	55.4%	12.3%	3.5%	3.3%
	V6	50.9%	8.8%	4.2%	3.9%
	Average	53.8%	12.0%	3.9%	5.7%

The percentage is calculated for each lead.

View Large Image Download to PowerPoint

Fig. 2. A, A clinical 10-second ECG signal on V2, where P waves were removed. B, ECG signal obtained by the addition of ventricle activity (VA) ECG signal (A) and AA signal (C). C, Original 10-second ECG simulated AA on V2. D, Estimated AA with the AF-Toolbox after applying the same high-pass filter. E, Original AA of amplified 4 times (C). F, Estimated AA of amplified 4 times (D); NMSE of 0.21 during the QRS complexes and of 0.17 during the JQ intervals.

Discussion 

Generally, when an algorithm is translated from research into industrial code, the performance of the algorithm suffers. In our case, the AF-Toolbox was found to outperform the other algorithm tested on the signals of the QRS intervals. The NMSE is only 0.75 ± 0.11, which is approximately 30% better than the best method tested in this study. The AF-Toolbox performed less well during the JQ intervals (outside the QRS complexes). However, the standard deviation is small, which indicates that the method used in the AF-Toolbox is more stable than the ABS and SB-EPFL methods. The AF-Toolbox is as stable as the ABS-EPFL (best method within this study on the signal part outside the QRS intervals). Finally, during the JQ intervals the AF-Toolbox performed as well as the SB-EPFL method (Table 2).

Therefore, we did not find the substantial performance loss initially feared for. The observed small loss in performance outside the QRS intervals is probably due to the higher noise floor caused by the integer arithmetic.

The overall performance of the AF-Toolbox is better than expected. The quality of the results is similar to the best one among the other algorithms tested despite being an industrially focused algorithm. The collaboration between the EPFL signal processing laboratory, in charge of the research part of the study and the biomedical research, and signal processing group from SCHILLER AG, performing the transition and implementation into a clinical device, has clearly been beneficial.

Conclusions 

We have shown that the AF-Toolbox algorithm, an industrial software application, can have a performance comparable to that found in previous studies. This was made possible by the close interaction between the 2 groups involved, each focusing on its specific expertise. We found that the proper signal separation of AA and VA is the key part of the procedure. When this separation performs correctly, the diagnostic part of the AF-Toolbox has great potential for clinical applications aimed at reliably detecting AF, which will elicit new approaches for its classification. The implemented visualization of the timing of the ventricular and atrial triggers provides a fresh look at the ECG. The AF-Toolbox will facilitate the search for meaningful signal features that may be used to identify the source of AF in individual patients.