Author + information
- Received April 24, 2018
- Revision received June 4, 2018
- Accepted July 3, 2018
- Published online August 15, 2018.
- Rahul G. Muthalaly, MD, MPHa,b,
- Raymond Y. Kwong, MD, MPHa,b,
- Roy M. John, MBBS, PhDa,b,
- Rob J. van der Geest, PhDc,
- Qian Tao, PhDc,
- Benjamin Schaeffer, MDa,b,
- Shinichi Tanigawa, MDa,b,
- Tomofumi Nakamura, MD, PhDa,b,
- Kyoichi Kaneko, MD, PhDa,b,
- Usha B. Tedrow, MD, MSa,b,
- William G. Stevenson, MDa,b,
- Laurence M. Epstein, MDa,b,
- Sunil Kapur, MDa,b,
- Paul C. Zei, MD, PhDa,b and
- Bruce A. Koplan, MD, MPHa,b,∗ ()
- aCardiovascular Division, Department of Medicine, Brigham and Women’s Hospital, Boston, Massachusetts
- bHarvard Medical School, Boston, Massachusetts
- cLeiden University Medical Center, Leiden, the Netherlands
- ↵∗Address for correspondence:
Dr. Bruce A. Koplan, Arrhythmia Service, Brigham and Women’s Hospital, 75 Francis Street, Boston Massachusetts 02115.
Objectives The aim of this study was to assess the utility of left ventricular (LV) entropy, a novel measure of myocardial heterogeneity, for predicting cardiovascular events in patients with dilated cardiomyopathy (DCM).
Background Current risk stratification for ventricular arrhythmia in patients with DCM is imprecise. LV entropy is a measure of myocardial heterogeneity derived from cardiac magnetic resonance imaging that assesses the probability distribution of pixel signal intensities in the LV myocardium.
Methods A registry-based cohort of primary prevention implantable cardioverter-defibrillator patients with DCM had their LV entropy, late gadolinium enhancement (LGE) presence, and LGE mass measured on cardiac magnetic resonance imaging. Patients were followed from implantable cardioverter-defibrillator placement for arrhythmic events (appropriate implantable cardioverter-defibrillator therapy, ventricular arrhythmia, or sudden cardiac death), end-stage heart failure events (cardiac death, transplantation, or ventricular assist device placement), and all-cause mortality.
Results One hundred thirty patients (mean age 55 years, 83% men, LV ejection fraction 29%, mean LV entropy 5.58 ± 0.72, LGE present in 57%) were followed for a median of 3.2 years. Eighteen (14.0%) experienced arrhythmic events, 17 (13.1%) experienced end-stage heart failure events, and 7 (5.4%) died. LV entropy provided substantial improvement of predictive ability when added to a model containing clinical variables and LGE mass (hazard ratio: 3.5; 95% confidence interval: 1.42 to 8.82; p = 0.007; net reclassification index = 0.345, p = 0.04). For end-stage heart failure events, LV entropy did not improve the model containing clinical variables and LGE mass (hazard ratio: 2.03; 95% confidence interval: 0.78 to 5.28; p = 0.14). Automated LV entropy measurement has excellent intraobserver (mean difference 0.04) and interobserver (mean difference 0.03) agreement.
Conclusions Automated LV entropy measurement is a novel marker for risk stratification toward ventricular arrhythmia in patients with DCM.
- implantable cardioverter-defibrillator
- magnetic resonance imaging (MRI)
- sudden cardiac death
Accurate sudden cardiac death risk stratification is important to optimize outcomes and minimize costs in clinical practice. Left ventricular ejection fraction (LVEF) has been the dominant risk stratification tool for sudden death risk in patients with dilated cardiomyopathy (DCM) to date. However, registries have demonstrated that LVEF is limited in its predictive accuracy and subject to significant fluctuations over time (1,2). Thus, alternative measures for arrhythmic risk stratification have been sought.
Myocardial replacement fibrosis detected by late gadolinium enhancement (LGE) imaging has been of interest in this regard. The presence and extent of LGE is associated with both arrhythmic and heart failure death risk in the DCM population (3,4). However, these methods are unsophisticated in that they do not characterize the whole gamut of pixels available in the ventricular myocardium. Rather, they set a threshold of signal intensity and assess the extent of pixels above that threshold.
We propose left ventricular (LV) entropy, a novel method of examining the myocardium in which the distribution of pixel intensities across the myocardium is assessed and compared. This method is based on a property used in photo editing, referred to as image entropy. Entropy in this context describes the complexity of an image; an image with completely homogenous pixels (e.g., a black square) would have entropy of zero. As the image becomes more complex, with many different pixel values, it has a higher entropy (5). Thus, by applying this principle to the LV myocardium, the “complexity” of the total myocardium can be quantified.
We hypothesized that primary prevention patients with DCM with higher LV entropy, and thus more myocardial heterogeneity, assessed by cardiac magnetic resonance imaging (CMR) would have a higher ventricular arrhythmia risk. Entropy is calculated as follows:where P(xi) is the probability distribution of signal intensities, x is signal intensity, and b is any chosen base (the software uses 2).
Study design and patient selection
Consecutive patients with DCM in a registry who presented for primary prevention implantable cardioverter-defibrillator (ICD) between 2006 and 2016 and underwent CMR within the 12 months prior to ICD were included. Patients were followed for outcomes from ICD implantation. DCM etiology was defined using clinical notes according to the European Society of Cardiology cardiomyopathy scheme (6). Patients’ primary prevention ICD indications were rated using the 2013 ICD appropriate use criteria (7). Covariate risk factors were assessed at the time of ICD implantation. All patients had ischemic cardiomyopathy excluded by either a functional test and/or coronary angiography. Patients were excluded if they met Felker’s criteria for ischemic cardiomyopathy on angiography or had positive functional test results that accounted for the cardiomyopathy (8). Additionally, all patients with endocardial LGE patterns on CMR consistent with infarction were excluded. Other exclusion criteria included infiltrative disease, hypertrophic cardiomyopathy, arrhythmogenic right ventricular cardiomyopathy, or a history of sustained ventricular tachyarrhythmia. The study was approved by the Institutional Review Board as part of the registry.
Patients were imaged using either a 3.0-T (Siemens Healthcare, Erlangen, Germany) or a 1.5-T (GE Healthcare, Chicago, Illinois) CMR scanner. All CMR images were acquired with electrocardiographic gating and breath-holding while patients were supine. Typical LGE CMR protocol parameters were a repetition time of 4.8 ms, an echo time of 2.3 ms, and an inversion time of 200 to 300 ms adjusted with the aim of nulling the anteroseptal segments of the left ventricle. LGE imaging started 10 to 15 min after a cumulative intravenous dose of 0.15 mmol/kg of gadopentetate dimeglumine (Magnevist, Bayer Pharmaceuticals, Leverkusen, Germany). Images were obtained in 8 to 14 matching short-axis (8 mm thick with 0 mm spacing) and 3 radial long-axis planes. Cine images in short and long axes were acquired in matching scan planes and typically with cine steady-state free precession imaging with the following parameters: a flip angle of 45° to 50°, in-plane resolution of 1.8 × 2.3 mm, and temporal resolution of 45 to 55 ms. All images were analyzed using specialized software (MedisSuite 3.0, Medis Medical Imaging Systems, Leiden, the Netherlands) by researchers blinded to outcomes.
The presence or absence of LGE was determined by a trained observer blinded to outcomes and was cross-checked against registry interpretations. Disagreements were resolved with a third observer. LGE mass quantification was performed by a blinded observer using the previously described full width at half maximum method (9). In brief, LV contours were manually traced on the LGE short-axis images, followed by delineation of remote normal and LGE regions. Artifact was manually excluded, and the software (MedisSuite 3.0) provided the final mass in grams. The full width at half maximum mass was defined as “dense scar.” LV entropy was semiautomatically determined by an observer blinded to outcomes. First, LV contours were traced on LGE short-axis images. The software then calculated the entropy of the LV myocardium using the formula described earlier on the basis of the signal intensity histogram. The software calculates the entropy by aggregating all slices of ventricular myocardium that are demarcated by the observer and drawing the entropy value from that distribution. Twenty random patients were selected to assess intraobserver and interobserver agreement for LV entropy with a minimum of 1-week interlude for intraobserver measurements. For interobserver agreement, a separate CMR-trained observer performed measurements blinded to outcomes and results of the original observer.
The primary outcome was arrhythmic events, defined as appropriate ICD therapies (including antitachycardia pacing or defibrillation for ventricular arrhythmias), sustained ventricular arrhythmia, or sudden cardiac death (according to consensus criteria) (10). Events were adjudicated by 2 electrophysiologists blinded to CMR data using clinical notes and device electrograms. Patients at our institution are followed with 3- to 6-month clinic visits in addition to transmitted ICD data. The 2 secondary outcomes were: 1) end-stage heart failure events (EHFE), including cardiac death, heart transplantation, or ventricular assist device placement; and 2) all-cause mortality. No patients with nonfatal secondary outcomes were censored from the primary outcome analysis. Secondary outcomes were determined using registry data, chart review, and the National Death Index. ICD complications were determined using chart review and classified into inappropriate shocks, infectious complications, and “other.”
Continuous data are presented as mean ± SD or median (interquartile range [IQR]) and were compared using Pearson’s correlation coefficient, Spearman’s rank coefficient, Student t tests, or analysis of variance. Categorical data are presented as percentages and frequencies and were compared using chi-square tests. To assess the difference between 1.5- and 3.0-T scans, we conducted a linear regression of entropy versus magnet strength, adjusting for age, body mass index (BMI), and ejection fraction in a larger cohort of 151 patients consisting of a mix of 3.0- and 1.5-T scans. This yielded a significant beta coefficient of −2.0 entropy for 1.5-T scanners compared with 3.0-T scanners. This constant was added to all entropy values from 1.5-T scanners to account for the difference. Inter- and intraobserver agreement was assessed using the Bland-Altman method. Collinearity was assessed using correlation and tolerances. Univariate hazard ratios were calculated using Cox proportional hazards regression. To avoid overfitting the Cox multivariate model because of a low event rate, we created a model probability score for each outcome by regressing variables with univariate p values <0.25 against LV entropy. Multivariate Cox regression models were generated by including the aforementioned model probability scores and LV entropy. Only univariate models were performed for mortality because of the low event rate. A p value <0.05 was deemed to indicate statistical significance. The proportional hazards assumption was tested using Schoenfeld residuals. The incremental predictive benefit of LV entropy for each outcome was assessed using the continuous net reclassification index for survival data, as described previously (11). Kaplan-Meier curves were compared using the log-rank test. Cut-point analysis using Liu’s method was used to determine an entropy value for generating survival curves (12). Stata version 14 (StataCorp, College Station, Texas) and R version 3.4 (R Foundation for Statistical Computing, Vienna, Austria) were used for analysis.
A total of 130 patients (mean age 54.8 ± 14.7 years, 108 men (83.1%), mean BMI 27.9 ± 6.7 kg/m2, and LVEF 29.4 ± 13.5%) were followed for a median of 3.2 years (IQR: 5.2 years). Other baseline demographics are shown in Table 1. Median time between CMR and ICD was 25 days (IQR: 117 days), and median ICD appropriate use rating was 9 of 9 (IQR: 1). A total of 101 patients were scanned with a 3.0-T scanner, and 29 patients were scanned with a 1.5-T scanner.
LV entropy and LGE
A total of 74 patients (56.9%) had LGE present, mean LV entropy was 5.58 ± 0.72, and median dense scar mass was 2.0 g (IQR: 5.96 g). The presence and extent of LGE were weakly associated with LV entropy (r = 0.20, p < 0.006 for dense scar mass). There was also a weak association for LV entropy with BMI (r = −0.40, p < 0.0001) and sex (5.5 in men vs. 5.8 in women, p = 0.02). Using Bland-Altman analysis, the mean intrarater LV entropy difference was 0.04, with limits of agreement of −0.18 to 0.26, as shown in Figure 1. The mean interrater LV entropy difference was 0.03, with limits of agreement of −0.39 to 0.33. Comparatively, the intrarater mean difference for scar mass was −1.12 g with limits of agreement of −11.68 to 9.44 g. The interrater mean difference for scar mass was −1.52 g, with limits of agreement of −17.74 to 14.71 g.
A total of 18 patients (14.0%) experienced arrhythmic events (13 ventricular tachycardia, 4 ventricular fibrillation, and 1 sudden cardiac death). The overall incidence was 4 events per 10 person-years. On univariate analysis, LV entropy (hazard ratio [HR]: 2.4; 95% confidence interval [CI]: 1.05 to 5.47; p = 0.04), LGE presence (HR: 4.5; 95% CI: 1.28 to 15.56; p = 0.02), dense scar mass (HR: 1.1/g; 95% CI: 1.06 to 1.18; p < 0.0001), diabetes (HR: 4.8; 95% CI: 1.81 to 12.71; p = 0.002), mineralocorticoid receptor antagonist use (HR: 2.8; 95% CI: 0.37 to 21.05; p = 0.02), and BMI (HR: 1.1; 95% CI: 1.04 to 1.15; p = 0.001) were significant predictors of the primary outcome, as shown in Table 2. An arrhythmia survival curve for LV entropy is shown in Figure 2.
On multivariate analysis for the arrhythmic outcome, using the model probability score that contained all variables with p values <0.25, LV entropy remained a significant predictor (HR: 3.5; 95% CI: 1.42 to 8.82; p = 0.007) and provided improvement in the continuous net reclassification index (0.345; 95% CI: 0.001 to 0.604; p = 0.04), as shown in Table 3. LV entropy remained significant when restricting the cohort to only patients with LVEFs <35% (HR: 2.9; 95% CI: 1.08 to 7.62; p = 0.03). Of 11 patients with LV entropy of <4.46, none experienced an arrhythmic event over a median of 2.4 years, despite 6 of these patients having scar. Examples of 2 patients, one with high entropy who experienced an arrhythmic event and cardiovascular death and the other with low entropy, are shown in Figure 3.
In an analysis of the 11 patients who experienced appropriate defibrillation, LV entropy maintained a strong adjusted HR for the outcome (HR: 3.9; 95% CI: 1.33 to 11.51; p = 0.01). In the subgroup of 101 patients who were scanned with a 3.0-T scanner, LV entropy was still associated with the primary outcome (HR: 3.1; 95% CI: 1.14 to 8.68; p = 0.03).
Seventeen patients (13.1%) experienced EHFE. Cardiac death occurred in 7 (5.4%), transplantation in 8 (6.2%), and ventricular assist device placement in 11 (8.5%). On univariate analysis, LV entropy (HR: 2.9), LVEF (HR: 0.94 per 1%), and New York Heart Association functional class (class II HR: 1.9; class III HR: 8.9; class IV HR: 9.1) were predictors of EHFE. Additional univariate predictors of EHFE are shown in Table 2.
On multivariate analysis including LV entropy with a model probability score containing all variables with p values<0.25 on univariate analysis, LV entropy was not a significant predictor of EHFE (HR: 2.03; 95% CI: 0.78 to 5.28; p = 0.14). Accordingly, LV entropy did not significantly improve the continuous net reclassification index (0.199; 95% CI: −0.252 to 0.436; p = 0.30). This remained the case when restricting the cohort to only patients with LVEFs <35%.
Seven patients (5.4%) experienced all-cause mortality. Univariate predictors of all-cause mortality were BMI (HR: 0.80) and New York Heart Association functional class (p = 0.02), as shown in Table 2.
Inappropriate ICD therapy occurred in 14 patients (10.9%), most commonly because of supraventricular tachyarrhythmias. ICD infection occurred in 4 patients (3.1%), and other ICD complications occurred in 7 (5.4%), most commonly lead related.
Novel markers that differentiate arrhythmic versus EHFE risk are of high potential impact in improving the use of ICDs in patients with DCM. In this pilot study, we have shown that LV entropy, a marker of myocardial heterogeneity, has novel prognostic association with arrhythmic but not EHFE in patients with DCM undergoing ICD for primary prevention of sudden cardiac death. Measurement of LV entropy is automated and easy to perform, and we have demonstrated its high measurement reproducibility. We believe these findings are intriguing and potentially offer a new method of assessment of arrhythmic risks from DCM.
Overall, our cohort experienced a similar incidence of arrhythmic endpoint as patients in the DANISH (Danish Study to Assess the Efficacy of Implantable Cardioverter Defibrillators [ICD] in Patients With Non-Ischemic Systolic Heart Failure on Mortality) trial (14% vs. 17.4%) (13). Our study design allowed ICD-based adjudication of arrhythmic events and constrained the population to primary prevention cases. In this context, LV entropy improved the prediction of arrhythmic events when added to a model that contained known clinical and CMR risk predictors. Furthermore, previous studies had demonstrated that LGE variables have incremental prognostic association with arrhythmic risk over LVEF (3,14). We observed that LV entropy added significant prognostic value for arrhythmic risks, over the clinical model including dense scar mass. We believe that our findings demonstrate the importance of tissue characterization using CMR in differentiating arrhythmic versus nonarrhythmic events but also highlight the inadequacy of current LVEF-based assessment of arrhythmic risk in primary prevention patients with DCM.
The association of scar mass and LV entropy with arrhythmic events is not surprising given the known anatomic basis for re-entrant arrhythmia circuits (15). LGE on CMR has been correlated with areas of slow conduction on electroanatomic mapping, corroborating this putative mechanism (16). Furthermore, scar mass has previously been associated with arrhythmic outcomes in nonischemic and ischemic cardiomyopathies (17,18). LV entropy provides a more sophisticated method of analyzing the distribution of LGE across the left ventricle, which appears to complement traditional LGE mass methods.
LV entropy displayed favorable reproducibility for both intraobserver and interobserver measurements. Thus, for clinical purposes, LV entropy may offer a realistic and robust method of risk assessment. Future studies may test LV entropy in conjunction with sophisticated functional measures of risk, such as global circumferential strain, to refine the prediction of events beyond traditional models (19).
LV entropy did not predict EHFE when added to a model probability score containing variables meeting the univariate threshold of p < 0.25. This characteristic, if verified in larger studies, may be helpful in assigning ICD therapy, as previous trials have shown that current markers predict not only arrhythmic death but also heart failure death, negating the protective benefit of an ICD for all-cause mortality (13). This question warrants examination through clinical trials, as is being done in the upcoming CMR-GUIDE (Cardiac Magnetic Resonance Guided Management of Mild-Moderate Left Ventricular Systolic Dysfunction) study (NCT01918215) and comparison alongside other modern CMR techniques such as T1 mapping and global circumferential strain (19,20).
First, this study was a single-center registry-based study and suffers from the biases commensurate with that design. Additionally, the study occurred at a tertiary center with easily accessible transplantation and ventricular assist device services, which may have inflated the EHFE rate. However, the event rate was still low, with only 18 arrhythmic events and 17 EHFE in total, which limits the complexity of models that can be built because of the risk for overfitting data and necessitates the use of model probability scores (21). Regarding external validity, it should be noted that this study included only patients with indications for primary prevention ICD, thus it may not accurately reflect all patients with DCM. Additionally, LV entropy does not examine the right ventricle, which may misclassify risk in cardiomyopathies that can affect the right ventricle preferentially, such as lamin A/C (22,23). We also used scanners with different static field strengths (1.5- and 3.0-T), using multiple linear regression with a larger sample of patients, adjusting for age and BMI, we found that the difference in entropy values was 2.00 greater with 3.0-T scanners. This point requires further exploration in future studies about what factors influence entropy values. Last, it must be acknowledged that the dependence of entropy on image acquisition time after contrast, sensitivity to suboptimal inversion times, and other imaging factors are unknown and need to be assessed. Future studies could also focus on histopathologic differences in hearts with different LV entropy values, how entropy correlates with other modern CMR markers such as T1 mapping, and understanding LV entropy in different cardiomyopathies.
LV entropy is a novel measure of heterogeneity, rapidly measured and highly reproducible. It demonstrates promising ability to predict arrhythmias in patients with DCM undergoing primary prevention ICD. If validated in larger multicenter studies, this feature of LV entropy may be valuable in stratifying arrhythmic risk for patients.
COMPETENCY IN MEDICAL KNOWLEDGE: CMR measures, such as the presence and mass of LGE on CMR, are well-established predictors of cardiovascular outcomes in patients with cardiomyopathy. However, measurement of LGE mass and presence is nonspecific for ventricular arrhythmia. The novel measure of LV entropy demonstrates additive diagnostic accuracy to LGE mass and clinical variables for arrhythmic events. LV entropy also had favorable reproducibility for interobserver and intraobserver measurements. An LV entropy value of <4.46 identified patients who experienced no ventricular arrhythmia during follow-up. Measurement of LV entropy may be a useful tool in the setting of DCM to predict cardiovascular outcomes.
TRANSLATIONAL OUTLOOK: Further studies documenting the utility of LV entropy for this purpose in both DCM and other cardiomyopathies are warranted to confirm these results and establish therapeutic thresholds for future trials. Work should also characterize the histopathologic correlates of LV entropy values in patients with DCM.
Dr. Muthalaly is supported by a Doctors-in-Training scholarship from Avant Mutual and an overseas research grant from Monash Health. Dr. John has received lecture honoraria from Abbott; and research support from Biosense Webster and Abbott. Dr. Stevenson is a coholder of a patent for needle ablation consigned to Brigham and Women’s Hospital. Dr. Epstein has received consulting fees from Boston Scientific, Medtronic, and Spectranetics. Dr. Tedrow has received consulting fees from St. Jude Medical; and research support from Boston Scientific and Biosense Webster. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose. Drs. Muthalaly and Kwong contributed equally to this work as first authors.
- Abbreviations and Acronyms
- body mass index
- confidence interval
- cardiac magnetic resonance imaging
- dilated cardiomyopathy
- end-stage heart failure event(s)
- hazard ratio
- implantable cardioverter-defibrillator
- interquartile range
- late gadolinium enhancement
- left ventricular
- left ventricular ejection fraction
- mineralocorticoid receptor antagonist
- Received April 24, 2018.
- Revision received June 4, 2018.
- Accepted July 3, 2018.
- Grimm W.,
- Christ M.,
- Bach J.,
- Muller H.H.,
- Maisch B.
- Halliday B.,
- Gulati A.,
- Ali A.,
- et al.
- Cover T.,
- Thomas J.
- Russo A.M.,
- Stainback R.F.,
- Bailey S.R.,
- et al.
- Felker G.M.,
- Shaw L.K.,
- O’Connor C.M.
- Piers S.R.,
- Tao Q.,
- van Huls van Taxis C.F.,
- Schalij M.J.,
- van der Geest R.J.,
- Zeppenfeld K.
- Hicks K.A.,
- Tcheng J.E.,
- Bozkurt B.,
- et al.
- Køber L.,
- Thune J.J.,
- Nielsen J.C.,
- et al.
- Neilan T.G.,
- Coelho-Filho O.R.,
- Danik S.B.,
- et al.
- Kazbanov I.V.,
- ten Tusscher K.H.W.J.,
- Panfilov A.V.
- Sasaki T.,
- Miller C.F.,
- Hansford R.,
- et al.
- Wu K.C.,
- Gerstenblith G.,
- Guallar E.,
- et al.
- Roes S.D.,
- Borleffs C.J.W.,
- van der Geest R.J.,
- et al.
- Mordi I.,
- Bezerra H.,
- Carrick D.,
- Tzemos N.
- Puntmann V.O.,
- Carr-White G.,
- Jabbour A.,
- et al.