Early Prediction of Heart Failure From Routine Cardiac CT Using Radiomic Phenotyping of Epicardial Fat

Figures

• 

  Figure 1. Radiomic Pipeline for EAT Profiling (A) Our end-to-end radiomic pipeline involves automated detection of the pericardial border, followed by extraction of the epicardial adipose tissue (EAT) and direct computation of 1,655 radiomic features. These, along with age and sex, are passed through a site-level harmonization process and then fed through a survival autoencoder. This architecture is designed to maximize the signal-to-noise ratio of radiomic signatures, while learning reproducible signatures linked to incident heart failure (HF). (B) The full algorithmic pipeline was trained in a development set of 7 centers (n = 59,327) and evaluated in an independent test set of 2 geographically distinct centers (n = 13,424) across the UK. The study population included individuals without known HF or myocardial infarction (MI) and with a 30-day blanking period. CCTA = coronary computed tomographic angiography; EAT = epicardial adipose tissue; FRPHF = fat radiomic phenotype for heart failure; HF = heart failure; MI = myocardial infarction; NHS = National Health Service; ORFAN = Oxford Risk Factors and Noninvasive Imaging cohort.

• 

  Figure 2. EAT Radiomic Profiling and Risk of New-Onset HF (A) Nelson-Aalen cumulative hazard curves for incident HF across discrete percentile groups (<10th, 10-25th, 25-50th, 50-75th, 75-90th, and >90th) of FRPHF score in the internal (dashed lines) vs external validation sets (solid lines). The curves present the observed (unadjusted) cumulative hazard and are accompanied by HRs (and 95% CIs) derived from multivariable Cox regression models adjusted for age, sex, and traditional cardiovascular risk factors, including epicardial adipose tissue volume, body mass index, and CAD-RADS categories. (B, C) Estimated cumulative incidence of new HF diagnosis at 1 (red), 3 (blue) and 5 (green) years after CCTA across EAT radiomic score percentiles (x-axis) in the internal and external validation cohorts (modeled for the median of each covariate and assuming no categoric risk factors). The shaded areas denote the 95% CIs. CAD-RADS = coronary artery disease–reporting and data system; other abbreviations as in Figure 1.

• 

  Figure 3. Discrimination of New-Onset HF at 5 Years According to FRPHF vs Traditional Risk Factors Plots summarizing the area under the time-dependent receiver operating characteristic curve (AUROC) for discriminating incident HF cases at 5-years in the (A) internal vs (B) external validation sets. Four main different models are presented: 1) a model consisting of age, sex, and EAT volume alone (model 1, red line); 2) a baseline model consisting of age, sex, traditional cardiovascular risk factors (body mass index EAT volume, diabetes mellitus, hypertension, history of ischemic heart disease, myocardial infarction, stroke, peripheral arterial disease, chronic kidney disease) and CAD-RADS categories (model 2, green line), 3) an FRPHF-only model (model 3, blue line), and 4) a composite model incorporating FRPHF alongside the baseline model predictors (model 4, orange line). P values represent the significance of pairwise comparisons against the baseline model (model 4 vs model 2). Abbreviations as in Figures 1 and 2.

• 

  Figure 4. Association of FRPHF With Incident HF Across Clinical Subgroups Forest plot presenting the age- and sex-adjusted hazard estimates for the association between FRPHF (in quartile [25-percentile] increments) and incident HF across the internal (blue) and external (red) validation sets in relevant clinical subgroups defined at the time of the CCTA scan. BMI = body mass index; VHD = valvular heart disease (includes aortic stenosis/regurgitation and/or mitral stenosis/regurgitation; severity unspecified); other abbreviations as in Figure 1.

• 

  Figure 5. Global Epicardial vs Focal Perivascular Radiomic Profiling (A) Representative axial slice illustrating the segmentation of the EAT depot used in the development of the FRPHF score, compared with a schematic visualization of the coronary perivascular adipose tissue (PVAT) surrounding the 3 main coronary arteries for development of the FRPPVAT score. (B) Correlation of FRPHF with FRPPVAT in the external validation cohort. (C, D) Age- and sex-adjusted association of (C) FRPHF and (D) FRPPVAT with HF events in the presence or absence of ischemic heart disease (IHD), as well as all-cause and cardiac mortality and acute myocardial infarction (MI). HU = Hounsfield units; other abbreviations as in Figure 1.

• 

  Figure 6. Example of Volumetric vs High-Throughput Radiomic Phenotyping of EAT (A) Example of EAT segmentation from an axial slice from a male patient, (B) with radiomic phenotyping by FRPHF revealing an elevated risk of incident heart failure with an FRPHF score in the 95th percentile. (C, D) Related example from a different case with radiomic EAT phenotyping revealing a more benign EAT remodeling profile with an FRPHF in the 39th percentile. HU = Hounsfield units; other abbreviations as in Figure 1.

• 

  Central Illustration. Radiomic Phenotyping of Epicardial Adiposity for Heart Failure Risk Stratification We present a radiomic fat phenotype (FRPHF) derived from epicardial adipose tissue on routine CCTA in the ORFAN multicenter cohort (9 sites, 72,751 scans). FRPHF captures adverse adipose remodeling linked to incident HF, demonstrates graded separation of cumulative HF incidence in internal and external validation, and provides calibrated incremental discrimination beyond traditional risk factors ,with consistent performance across subgroups and imaging protocols. CCTA = coronary computed tomographic angiography; EAT = epicardial adipose tissue; FRPHF = fat radiomic phenotype for heart failure; HF = heart failure.