Computers in Cardiology Challenge

Available online at www.sciencedirect.com

Journal of Electrocardiology 41 (2008) 630 – 635 www.jecgonline.com

Using inverse electrocardiography to image myocardial infarction—reflecting on the 2007 PhysioNet/Computers in Cardiology Challenge Fady Dawoud, BENG, a,⁎ Galen S. Wagner, MD, b George Moody, PhD, c B. Milan Horáček, PhD a a

School of Biomedical Engineering, Dalhousie University, Halifax, NS, Canada b Duke University Medical Center, Durham, NC, USA c Harvard-MIT Division of Health Sciences and Technology, Cambridge, MA, USA Received 10 June 2008; revised 18 July 2008; accepted 18 July 2008

Abstract

The goal of the 2007 PhysioNet/Computers in Cardiology Challenge was to try to establish how well it is possible to characterize the location and extent of old myocardial infarcts using electrocardiographic evidence supplemented by anatomical imaging information. A brief overview of the challenge and how different challengers approached the competition is provided, followed by detailed response of the first author to integrate electrophysiologic and anatomical data. The first author used the provided 120-electrode body-surface potential mapping data and magnetic resonance imaging heart and torso images to calculate epicardial potentials on customized ventricular geometries. A method was developed to define the location and extent of scar tissue based on the morphology of computed epicardial electrograms. Negative Q-wave deflection followed by R-wave on the left ventricular surface seemed to correspond with the location of the scar as determined by the gadolinium-enhanced magnetic resonance imaging gold standard in the supplied data sets. The method shows promising results as a noninvasive imaging tool to quantitatively characterize chronic infarcts and warrants further investigation on a larger patient data set. © 2008 Elsevier Inc. All rights reserved.

Keywords:

Inverse electrocardiography; Myocardial infarction; PhysioNet; Computers in Cardiology Challenge 2007

Introduction The idea of the 2007 PhysioNet/Computers in Cardiology Challenge was to bring together the disciplines of electrocardiography and myocardial imaging in diagnosing myocardial infarction. The goal was to try to establish how well it is possible to characterize the location and extent of moderate to large and relatively compact old (healed) myocardial infarcts using electrocardiographic evidence supplemented by anatomical imaging information.1 The topic of the challenge was proposed by Galen Wagner (Duke University, Durham, NC) and was developed as a result of collaboration among multiple groups and investigators (George Moody, MIT and PhysioNet; the VVRED [Virtual Visual Reconstructed Electrocardiographic Display] group; ⁎ Corresponding author. E-mail address: [email protected] 0022-0736/$ – see front matter © 2008 Elsevier Inc. All rights reserved. doi:10.1016/j.jelectrocard.2008.07.022

and the MALT [Magnetic and Electrocardiographic Technologies] group). The data provided for the challenge was chosen from the MALT study2 patient population (2 training cases and 2 test cases: each case with old [1 year] infarct). The body-surface potential mapping (BSPM) data, consisting of electrocardiogram (ECG) data for 352 torso-surface sites, were provided for a single averaged PQRST complex sampled at 1 kHz for all the 4 cases. The data also contained the standard 12 leads, the 7 unweighted Frank leads, and the Frank orthogonal XYZ leads. Selected anatomical 5-mm magnetic resonance imaging (MRI) transaxial images were available for all 4 cases, whereas gadolinium-enhanced transaxial MRI (GEMRI) images showing the infarction substrate were only provided for the 2 training cases and were not provided to challengers for the 2 test cases. Transaxial GE-MRI provided a reference (the gold standard) for scoring submissions of participants in the

F. Dawoud et al. / Journal of Electrocardiology 41 (2008) 630–635

631

Fig. 1. 17 segmentation of the left ventricle, adapted from Ref. [4].

challenge. Analysis of the GE-MRI data to define the reference was done by the MALT investigators using custom developed software that, semiautomatically, subdivided the left ventricular (LV) myocardium into 17 segments and delineated hyperenhanced myocardium as infarct, thus, transforming infarct location and extent in terms of the LV 17-segment representation.3 Scoring of submissions was based on 3 criteria: EPD, percentage discrepancy between the extent of the infarct as estimated and as determined from the gold standard; SO, overlap between the sets of infarct segments as estimated and as determined from the gold standard; and CED, distance between the centroid of the infarct as estimated and as determined from the gold standard. The 17-segment representation of the LV4 was used to report all results (Fig. 1). The first author of the manuscript entered the challenge (without access to the gold standard), and his methods and results are described in detail in the next 2 sections. The methods and results obtained by the other 5 teams of participants are described briefly in the discussion section, which includes a comparison of all 6 approaches. Methods

model. The supplied body-surface ECGs, which were recorded at 120 anatomical locations on the torso according to the standard Dalhousie configuration, were interpolated to the 352 locations corresponding to the standard Dalhousie torso projected on the new customized torso geometry. The BSPM data contained one averaged PQRST complex sampled at 1000 Hz. The model used to solve the inverse problem is explained in detail by Horáček and Clements.5 Briefly, the boundary element method is used to model the human torso (as a homogeneous and isotropic volume conductor) and the epicardial ventricular surface (as the region containing all bioelectric sources). The relationship between electric potential on epicardial ventricular elements and the torso elements can be represented by the equation ΦB = AΦH, where ΦB is the body-surface potential, A is the transfer coefficient matrix incorporating geometry and conductivity properties of the volume conductor model, and ΦH the epicardial potential. The ill-conditioned property of the transfer matrix A requires using a regularization scheme that tries to minimize a similar problem to get a smoother and more stable solution for ΦH. Eq. (1) describes the minimization problem in the case of Tikhonov regularization6: n o min jjAUH  UB jj2 þ kjjBUH jj2 ð1Þ

The first author analyzed the supplied MRI images (DICOM format) of the 4 cases (which did not include the GE-MRI gold standard data) with Amira 4.1 software (Mercury Computer Systems, Chelmsford, MA) to create customized surfaces of the ventricles and torso geometries for use in the inverse procedure. Discretized ventricular surfaces in the 4 cases consisted of 1000, 700 (Fig. 2), 600, and 500 triangular elements. Transfer coefficients relating potentials on the ventricular epicardium to measured body-surface potentials were calculated with constant interpolating function on triangular elements assuming a homogenous torso

where λ is the regularization parameter that controls the amount of smoothing and B is a geometry matrix related to the ventricular surface (identity matrix in zero-order regularization and surface Laplacian in second regularization). The epicardial potential ΦH was calculated using Tikhonov second-order regularization method, and the regularization parameter was obtained according to the Lcurve method.7 A method was developed to group inverse epicardial electrograms based on similar morphology. Electrograms were classified into 6 categories: QR, qR, Qr,

632

F. Dawoud et al. / Journal of Electrocardiology 41 (2008) 630–635

Fig. 2. Frontal (right) and left (left) sagittal views of customized torso and epicardial surfaces for case 2. Each surface consists of 700 triangular elements.

RS, rS, and Rs according to the usual nomenclature where Q/ q represents early negative deflection, R/r represents positive deflection, and S/s represents late negative deflection. Capitalization denotes relative magnitude of the two major deflections in the inversely calculated electrocardiogram based on a threshold ratio chosen to be 0.25. Custom-written MATLAB (Mathworks Inc, Natick, MA) routines were used for all data processing and analysis.

interventricular groove. Electrogram morphologies on the LV appear to have inverted mirror images on the RV, that is, qR (red) on LV reflects as rS (light blue) on RV, and Rs (black) on LV reflects as Qr (yellow) on RV. Based on results obtained from the training cases 1 and 2, criteria for determining both the location and centroid of the underlying infarction were based on the presence of major Q-wave morphology electrograms. Extent of the infarction was qualitatively estimated by averaging weights

Results Fig. 3 shows color-coded distributions of inverse electrogram morphologies for training case 2, displayed on the customized heart geometry. Using GE-MRI analysis, the infarct region was determined to be segments 3, 4, 9, and 10. The inferior outline of the infarct region as determined from the GE-MRI is shown in gray (segments 3, 4, 9, and 10). Basal parts of the inferior surface of both the right ventricular (RV) and LV show QR morphology (blue), whereas the inferior LV midcavity region shows qR morphology (red). It was noticed that infarcted myocardium extends over areas with a major early negative deflection (largely on QR, Qr regions and sometimes on qR). In case 2, a large portion of the infarct occurs on areas showing QR morphology. Presence of QR-type electrograms on basal RV regions could be reflecting infarcted myocardium extending over inferoseptal LV (segments 3 and 9). Fig. 4 shows distribution of inverse electrogram morphologies for training case 1, where the infarct was located primarily in the septum (segments 1, 2, 3, 8, 9, 13, 14, and 15). A region of QR pattern is seen overlaying the anterosuperior interventricular groove, extending slightly towards the LV and RV. On the inferior surface, a reciprocal morphology of RS deflection (green) overlays the anterior

Fig. 3. Anterosuperior (top left) and inferior views (bottom left) of case 2 (training) epicardial ventricular surface. The color coding is according to one of the 6 electrogram types (QR, RS, qR, rS, Qr, and Rs) that correspond to blue, green, red, cyan, yellow, and gray, respectively. The QRS complex morphology coding is according to the conventional nomenclature: Q represents an early negative deflection, 5 represents a late negative deflection and R represents an early positive deflection and capitalization represents the relative magnitude of the two peaks in each electrogram. The average electrogram for each of the 6 electrogram types is plotted (right).

F. Dawoud et al. / Journal of Electrocardiology 41 (2008) 630–635

Fig. 4. Case 1 (training) electrogram morphology results. Format is the same as in Fig. 3.

assigned to segments. The weights were chosen based on the spread of estimated infarct in the segment and relative size of the segment. Fig. 5 show results obtained using the morphology classification method for test case 3. A region of QR morphology overlays basal as well as midcavity inferior/ inferolateral parts of the LV. Extension of QR-type electrograms over the inferior RV base suggests (as seen in case 2 results) basal inferoseptal spread of the infarct. Segments estimated to contain infarct were 3, 4, 5, 10, and 11 (scored 0.556 of 1) with centroid located in segment 4 (one segment away) extending over 35% of LV mass (departing 17% from gold standard reference). Results for case 4 are shown in Fig. 6. Region of QR morphology appears on basal RV and LV extending toward anterolateral LV. qR morphology is seen on the remainder of the inferior surface, similar to case 2. Estimated infarcted segments were 3, 4, 5, 6, 9, 10, and 11 (scored 0.333 of 1), with centroid located in segment 4 (2 segments away) extending

633

Fig. 6. Case 4 (test) electrogram morphologies. Format is the same as in Fig. 3.

over 40% of LV mass (departing 26% from gold standard reference). Table 1 summarizes the results as estimated by the morphology classification inverse solution method and the gold standard reference interpretation made available after the end of the challenge. Discussion Reflecting on the challenge The 2007 PhysioNet/Computers in Cardiology Challenge was unique in giving participants the opportunity to approach the challenge at different levels of modeling (cellular automaton, inverse solution, body-surface mapping, 12-lead ECG, and vectorcardiography). The challenge has attracted a relatively small number of entrants compared to previous years because of its multi-disciplinary nature that requires combining signal processing, image processing, and modeling techniques. Nevertheless, by having analyzed the same data sets, the participants have made possible an objective comparison of the variety of approaches they have used in the challenge, thus, also making it possible to identify promising approaches as subjects for further study. Table 1 Results of the inverse solution morphology method and the reference information Scoring criterion Extent of infarct GE-MRI reference Inverse estimate Segments of infract GE-MRI reference

Inverse estimate

Fig. 5. Case 3 (test) electrogram morphologies. Format is the same as in Fig. 3.

Case 1

Case 2

Case 3

Case 4

31% 25%

30% 35%

52% 35%

14% 40%

1, 2, 3, 8, 9, 13, 14, 15 2, 3, 8, 9, 14

3, 4, 9, 10

3, 4, 5, 9, 10, 11, 12, 15, 16 3, 4, 5, 10, 11

1, 9, 10, 11, 15, 17 3, 4, 5, 6, 9, 10, 11

10 or 11

15

4

4

Centroid of infarct GE-MRI reference

8

Inverse estimate

9

3, 4, 5, 9, 10, 11 3 or 4 or 9 or 10 10

634

F. Dawoud et al. / Journal of Electrocardiology 41 (2008) 630–635 8

Challenger 1 (Farina et al ) used a cellular automaton model of the heart taken from the Visible Male Project. The model contained ventricles, atria, SA node, AV node, and Purkinje network with 1-mm3 cell size and 200k active cells (endo-, epi-, and midmyocardial transmembrane potential templates). The thorax model was a heterogeneous volume conductor defined on a finite element mesh. The location of infarct was determined by an optimization procedure of the electrophysiologic heart model with an introduced infarction scar to best fit the multi-channel BSPM measured on the surface of the thorax. Challenger 2 (Dawoud9) used the method of inverse electrocardiography to calculate epicardial potentials from body-surface potentials using a customized homogenous torso and ventricular surfaces obtained from the supplied MRI images, as detailed above. Challenger 3 (Zarychta et al10) proposed to use metrics derived from body-surface potential maps, namely, peak R-wave amplitude, ST-segment level, peak T-wave amplitude, and QTinterval, to try to approximate location of the scar. Challenger 4 (Mneimneh and Povinelli11) used an iterative Gaussian mixture model approach applied to ordinary 12lead ECG to identify the infarcted segments. The 2 challenge training data sets were augmented with the PTB Diagnostic ECG augmented with the PTB Diagnostic ECG database (http://physionet.org/physiobank/database/ptbdb/) as a training set to identify the infarcted segments. Challenger 5 (Ghasemi et al12) used a method based on vectorcardiography that assumes the heart vector to be proportional to relevant active depolarization area and derives location and extent of scar based on this quantity. Challenger 6 (SadAbadi et al13) applied a method to relate certain body-surface potential features, namely, Q-wave amplitude and ST dispersion, at each torso node to estimate infarcted myocardium. Table 2 summarizes the results obtained by all 6 teams of participants. If we try to observe some trends from the small sample of the submitted responses, we notice that the models that did not account for anatomical or electrophysiologic details achieved better scores. Challengers 4, 5, and 6 who used electrocardiographic information only derived from 3 leads, 12 leads, or body-surface nodes shared top ranks in the 3 scoring categories. On the other hand, the more involved models of challengers 1 and 2 did not perform as well. This could be because analysis of the limited-lead data set is well understood and easily applicable, whereas, on the other hand, the more complex models are more sensitive to

modeling parameters and are less studied. Another observation to note is that the first place winner, challenger 4, was the only participant who used a machine learning algorithm supplanted by a larger 12-lead data set to better train the learning algorithm. This could indicate the role that artificial intelligence algorithms guided by large learning databases can play to further improve performance. Follow-up studies should take into consideration development of larger data sets that include control data free of infarct as well as pathologic data with myocardial infarction (MI) located in various LV segments. Also, a unified representation of data at different levels (MRI images, discretized surfaces, segmentation of ventricular surfaces, and MI exact location) is useful to allow more quantitative comparison of different methods. The inverse ECG morphology method The inversely calculated epicardial potentials using BSPM data and a transfer matrix constructed from customized torso and heart geometries provide a rational way to image local epicardial activity based on a physiologic approach. Displaying regions of similar electrogram morphology would appear to be a useful way to summarize progression of electrical activation in one map. The appearance of Q-wave–type electrograms (Qr, QR, and qR) on the LV was observed to correspond to infarcted areas. This result is consistent with clinical presentation of abnormally deep and wide Q-waves in certain 12-lead ECG leads in cases of old MI. It is interesting to note that in case 4, which did not present any abnormal Q-waves in the standard 12-lead ECG, the inversely calculated heart electrograms still indicated areas with QR morphology over LV inferobasal regions. The extent of the infarct as estimated by the in verse electrogram morphology method tended to overestimate the infarcted region. This can be overcome by adding additional amplitude, slope, integral, or duration constraints to the currently simple criterion of relative peak amplitude of 0.25 used to classify electrograms. Other methods were attempted (results not shown) to infer location of infarct with different degrees of success. These include QRS integral, peak positive repolarization amplitude, steepest negative slope value during depolarization, and J-point elevation. These criteria would need to be tested more systematically on a large number of cases to evaluate their effectiveness and determine appropriate threshold values for infarcted tissue.

Table 2 PhysioNet/Computers in Cardiology Challenge 2007 final results for all participants Entrant (entries) 1 (2) 2 (3) 3 (1) 4 (2) 5 (2) 6 (3)

EPD

SO

CED

Case 3

Case 4

Total

Rank

Case 3

Case 4

Total

Rank

Case 3

Case 4

Total

Rank

43 17 22 25 11 2

14 26 17 2 21 6

57 43 39 27 32 8

6 5 4 2 3 1

0.400 0.556 0.444 0.900 0.600 0.500

0.167 0.300 0.000 0.250 0.333 0.444

0.567 0.856 0.444 1.150 0.933 0.944

5 4 6 1 3 2

1 1 1 0 0 0

1 2 4 1 1 1

2 3 5 1 1 1

4 5 6 1 1 1

Adapted from Ref. [1].

Overall rank 5 4 6 1 3 1

F. Dawoud et al. / Journal of Electrocardiology 41 (2008) 630–635

There are some limitations for applying the inverse solution method, particularly on a large scale. Most important is the sensitivity of the inversion procedure to geometrical errors introduced by inaccurate orientation of the ventricular surface and uncertainty about body-surface leads locations.14 Also, manually constructing accurate geometries is time consuming and entails additional cost of using scans from a noninvasive imaging modality. Another limitation relates to tendency of the inverse procedure (assuming a homogeneous torso) to preserve electrogram amplitudes less accurately than morphology, which prompted our approach to use a patternoriented algorithm to classify electrograms. Also, the inverse model calculates electrograms on the epicardial ventricular surface and thus activity on the interventricular septum is largely masked by the activity of the RV wall. In conclusion, our method of approximating infarct location and extent using an inverse procedure with customized torso and heart geometries as applied to the 2007 PhysioNet/Computers in Cardiology challenge shows promise and warrants further validation on a larger number of both control cases (free from MI) to establish normal electrogram patterns and distribution, as well as on MI cases to investigate variability, evaluate accuracy, and improve detection algorithms. For more information about the 17 segments, contact Engblom. Acknowledgments

2.

3.

4.

5.

6. 7. 8. 9.

10.

11.

12.

This study was supported in part by grants from the Canadian Institutes of Health Research.

13.

References

14.

1. PhysioNet. PhysioNet/Computers in Cardiology Challenge 2007: electrocardiographic imaging of myocardial infarction. [2007

635

September 21] http://physionet.org/challenge/2007/ [Accessed on 27 May 2008]. Engblom H, Foster JE, Martin TN, et al. The relationship between electrical axis by 12-lead electrocardiogram and anatomical axis of the heart by cardiac magnetic resonance in healthy subjects. Am Heart J 2005;150:507. Heiberg E, et al. Semi-automatic quantification of myocardial infarction from delayed contrast enhanced magnetic resonance imaging. Scand Cardiovas J 2005;39:267. Cerqueira MD, et al. Standardized myocardial segmentation and nomenclature for tomographic imaging of the heart: a statement for healthcare professionals from the Cardiac Imaging Committee of the Council on Clinical Cardiology of the American Heart Association. Circulation 2002;105:539. Horáček BM, Clements JC. The inverse problem of electrocardiography: a solution in terms of single- and double-layer sources on the epicardial surface. Math Biosci 1997;144:119. Tikhonov AN, Arsenin VY. Solution of ill-posed problems. New York: Wiley; 1977. Hansen PC. Analysis of discrete ill-posed problems by means of the L-curve. SIAM Review 1992;34:561. Farina D, Dossel O. Model-based approach to the localization of infarction. Proceedings of the 34th Computers in Cardiology Conference; 2007. Dawoud F. Using inverse electrocardiography to image myocardial infarction. Proceedings of the 34th Computers in Cardiology Conference; 2007. Zarychta P, Smith FE, King ST, et al. Body surface potential mapping for detection of myocardial infarct sites. Proceedings of the 34th Computers in Cardiology Conference; 2007. Mneimneh MA, Povinelli RJ. RPS/GMM approach toward the localization of myocardial infarction. Proceedings of the 34th Computers in Cardiology Conference; 2007. Ghasemi M, Jalali A, SadAbadi H, et al. Electrocardiographic imaging of myocardial infarction using heart vector analysis. Proceedings of the 34th Computers in Cardiology Conference; 2007. SadAbadi H, Jalali A, Ghasemi M, et al. Variation of ECG features on torso plane: an innovative approach to myocardial infarction detection. Proceedings of the 34th Computers in Cardiology Conference; 2007. Ramanathan C, Rudy Y. Electrocardiographic Imaging: II. Effect of torso inhomogeneities on noninvasive reconstruction of epicardial potentials, electrograms, and isochrones. J Cardiovascr Electrophysiol 2001;12:241.