More Than Meets the Eye

Sudden cardiac death (SCD) from ventricular arrhythmia (VA) remains a devastating event with low survival rates for out-of-hospital arrests. Implantable cardioverter defibrillators (ICDs) reliably prevent SCD resulting from VA. Targeting placement based on the current left ventricular ejection fraction (LVEF) criterion of ≤35% has merits on the population level but suffers when applied to the individual patient.¹ Such an approach results in over-treatment of many patients who will never receive an appropriate device discharge yet are at risk for shortterm and long-term device complications and inappropriate firings. By contrast, many people with preserved LVEF, who constitute the largest absolute number of arrest victims, are unprotected and may succumb to SCD as their first manifestation of heart disease.

Decades of research have highlighted the critical importance of improving SCD risk stratification at the individual level by identifying additional factors for increased risk beyond LVEF. A key predictor is the status of the underlying myocardial substrate. Structural remodeling from both ischemic scar and nonischemic fibrosis results in regions of inhomogeneous tissue conduction that sustain re-entrant ventricular arrhythmias. Cardiovascular magnetic resonance with late gadolinium enhancement (CMR-LGE) has proved to be the noninvasive gold standard for quantifying and characterizing myocardial scar. Many studies have validated the association between increased VA risk and larger total amounts of scar as well as more extensive heterogeneous regions of scar, as defined by intermediate levels of signal intensity (gray zone) reflecting the border region of admixed normal myocytes and fibrotic bundles.^2,3 However, LGE techniques are predicated on visible differences in signal intensity and rely on categorical thresholds to distinguish normal from densely or heterogeneously scarred tissue.

The signal intensity profile and distribution assessed as a continuous variable across the entire left ventricle (LV) could yield richer information about scar architecture beyond that of LGE. Radiomics is an evolving approach that involves the extraction of distinctive quantitative features from clinical imaging to allow inferences about tissue composition.⁴ Texture analysis, a component of radiomics, assesses variations in the gray scale signal intensities within an image, indistinguishable by the naked eye. Entropy is a texture property that describes the regularity of the signal intensity patterns within an image. Low entropy signifies a predictable and compact pattern of pixel intensities reflecting a uniform homogeneous myocardial composition. By contrast, high entropy denotes an irregular broad range of regional signal intensity values, a surrogate reflecting a mixture of tissue types. Among recent studies of patients with ischemic and nonischemic cardiomyopathy, scar and/or total LV entropy was associated with increased risk for VA events.^5-8

In this issue of JACC: Cardiovascular Imaging, Antiochos et al⁹ enrich the current literature with their study investigating the association between total LV entropy and (MACE) (all-cause mortality, incident sustained ventricular arrhythmia, or heart failure hospitalization). Their single-center cohort comprised 583 patients with and without ischemic heart disease presenting for clinical evaluation of nonsustained ventricular tachycardia (NSVT) (54%), sustained VT, or aborted SCD who underwent CMR-LGE. Most patients were male (61%) with a low prevalence of comorbidities (11% prior heart failure, 7% prior myocardial infarction). Overall, LVEF was normal (54% ± 13%), as were right ventricular ejection fraction and biventricular volumes and masses. The prevalence of LGE was 39%, with a low overall scar extent (3.7% ± 6.8% of LV mass) and a predominantly subendocardial scar pattern (48%). Total LV entropy significantly differed by MACE occurrence and was a strong univariable predictor (HR: 1.88 [95% CI: 1.63-2.17] per SD, P < 0.001) with a higher chi-square value compared with LGE extent (HR: 1.91 [95% CI: 1.61-2.28] per 10% of LV mass, P < 0.001). After multivariable adjustment, LV entropy remained strongly associated with MACE (adjusted HR: 1.61 [95% CI: 1.32-1.96], P < 0.001) and incrementally predictive compared with LGE extent, which lost significance, likely reflecting the observed correlation between the 2 metrics (r = 0.62, P < 0.001). Notably, LVEF remained a strong predictor of MACE in multivariable modeling. Overall, adding LV entropy improved the baseline model C-statistic from 0.725 to 0.754 (P = 0.003).

Post-hoc subgroup analyses provided novel insights. Higher LV entropy was associated with a higher incidence of each component of the composite MACE outcome. It was most substantial for heart failure hospitalization and similar for sustained VT and all-cause mortality (unadjusted HR per SD of 2.50 vs 1.80 and 1.83, respectively). As expected, the incidence of MACE was higher among patients presenting with sustained VT or aborted SCD versus NSVT. However, LV entropy remained a significant predictor among both groups but was more robust in the NSVT subgroup (HR: 2.16 vs 1.43). For the outcome of incident VT, LV entropy was a stronger predictor among NSVT patients than among sustained VT patients (HR: 2.43 vs 1.34). Similarly, LV entropy predicted incident VT regardless of LGE presence, but the contribution was stronger among those without LGE (adjusted HR: 2.58 vs 1.30).

Although the study highlights the potential value of LV entropy as a marker of tissue heterogeneity that is more automated and simpler to derive than threshold-based gray zone analysis techniques, several limitations warrant acknowledgment. Whereas LV entropy likely incorporates aspects of total scar and gray zone extent, the biological context needs further validation. The time course of changes in LV entropy as the LV remodels, both positively and negatively, both spontaneously and in response to medical therapy, will affect subsequent risk, and such serial changes remain to be investigated. Although the current study acknowledged and accounted for differences in LV entropy measurement by scanner field strength, other technical factors affecting repeatability and reproducibility remain unknown. Entropy is only 1 feature of texture, and the current approach does not incorporate 3D spatial complexity as can be probed by other methods.¹⁰ More extensive multiparameter radiomics open the possibility of extracting and integrating a plethora of imaging features from multiple sequences, including cines and T1 mapping, as does more sophisticated deep learning of imaging features for improved VA risk prediction.¹¹

The comprehensive analysis of the current study cohort nonetheless extends the existing knowledge base by emphasizing the increased risk for adverse outcomes, including for VA, even among those presenting with nonsustained VT, preserved LVEF, and absence of LGE. Whereas LVEF and LGE scar metrics pose clear risk factors for adverse outcomes, there is the opportunity to improve individualized risk prediction by additional stepwise integration of automated metrics, such as LV entropy, that offer a glimmer of the rich structural and tissue characterization data embedded in CMR that likely further distinguish patients but remain unexploited. Ultimately, however, the field needs to advance from the many provocative observational studies such as the current study to randomized controlled trials that put to the test a CMR-driven strategy for prophylactic ICD insertion and thereby prove definitively whether we can replace the decades-old LVEF threshold approach.