Introduction

Pericardial adipose tissue (PAT) is the visceral adipose tissue compartment surrounding the heart and coronary vasculature. An increasing body of evidence highlights associations between greater amounts of PAT and poorer cardiovascular outcomes [1,2,3,4,5]. Furthermore, higher PAT has been linked to adverse cardiovascular phenotypes, independent of multiple other measures of adiposity [6]. These associations are highly suggestive of a distinct mechanistic role of PAT in driving adverse cardiac remodelling, which are precursors of heart failure [6].

The mechanisms through which PAT influences myocardial structure and function are likely multifactorial, involving paracrine, vasocrine, and inflammatory pathways. PAT is known to secrete inflammatory factors and lipid metabolites [7], and this metabolic and secretory activity has been highlighted as an important factor driving adverse cardiovascular outcomes. At a cellular level, the secretome of PAT has been shown to adversely impact cardiomyocyte contractility [8], metabolism [9], and disrupt adhesion molecule expression in cardiac endothelial cells [10]. In the setting of ischemic heart disease, patterns of coronary atherosclerosis have been shown to closely follow superficial PAT distribution [

Material and methods

We provide a schematic illustration of the study pipeline in Fig. 1.

Fig. 1
figure 1

Schematic illustration of the study pipeline. A The segmentation of pericardial adipose tissue (PAT); color overlay represents the segmentation results as derived by Bard et al. B The radiomics feature extraction: PAT segmentation output was used to extract radiomics features from cardiovascular magnetic resonance imaging data. C The final cohort, which was assembled from prevalent and incident heart failure cases and randomly selected control individuals from the UK Biobank dataset

Full size image

Setting and study population

The UK Biobank is a cohort study incorporating more than 500,000 individuals from across the UK, aged 40–69 years old at recruitment between 2006 and 2010. Baseline assessment included socio-demographics, lifestyle, environmental factors, medical history, and a range of physical measures. Extensive electronic health record (EHR) linkages permit prospective tracking of health outcomes for all participants. The UK Biobank Imaging Study, which includes CMR, aims to scan a randomly selected 20% subset of the original participants.

Ascertainment of heart failure status

HF status was ascertained using diseases codes from UK Biobank assessments and linked EHRs (Supplementary Table 1). Prevalent HF was defined as HF present at the time of imaging. Incident HF was considered as first occurrence of HF after imaging. The censor date was 30 September 2021 for incident HF outcomes giving an average follow-up of 3.7 ± 1.5 years from imaging.

Definition of the comparator group

Participants with CMR available and no record of HF (prevalent or incident) were eligible for inclusion in the control group (n = 42,327). There was substantial imbalance between cases and eligible controls. Such imbalance results in poor model performance as the model prediction will be dominated by the majority class [16]. Given the extreme imbalance in our dataset, with a large number of controls compared to cases, and considering computational constraints we applied random undersampling to reduce the frequency of the controls relative to the cases. This approach randomly removes subjects from the majority classes to reach a final set of subjects in the majority class that are similar to the minority class. The final sample contained an equal number of randomly selected non-HF controls for both the prevalent and incident HF groups.

Characterising the study sample

We accessed self-reported fields for participants’ educational level and smoking status. Material deprivation is reported as the Townsend index. Physical activity was measured via self-reported responses to the International Physical Activity Questionnaire. Diabetes, hypertension, and high cholesterol status were ascertained using information from self-report questions, physical measurements, and EHR data (Supplementary Table 2).

Image acquisition

CMR scans were performed according to a pre-defined acquisition protocol using 1.5-Tesla scanners (MAGNETOM Aera, Syngo Platform VD13A, Siemens Healthcare) [17]. Cardiac function was assessed using standard long- and short-axis balanced steady-state free precession cine sequences.

Extraction of pericardial fat segmentations

PAT segmentation was performed using an automated quality-controlled pipeline developed and validated in the UK Biobank and in an external cohort, as described by Bard et al [15]. In brief, PAT was measured from standard four-chamber cine images (single 2D slice) at phase 1 of the image cycle, which approximates end-diastole (ED). The contour was drawn to select areas of high signal intensity bordering the epicardial surface of the left and right ventricular myocardia. The ground truth manual segmentation was based on a sample of 500 randomly selected UK Biobank imaging sub-study participants using CVI42 post-processing software (Version 5.11, Circle Cardiovascular Imaging Inc.). Using the manual segmentation, a MultiResUNet neural network with Bayesian modification was trained for automated PAT segmentation with inbuilt quality control. Overall, the performance of the algorithm in test set relative to manual segmentations was good and very similar to the agreement between human observers (mean Dice score = 0.8) [15]. Automated PAT analysis was performed for all participants with adequate CMR imaging available (n = 42,929).

Background to radiomics

Radiomics is an analysis technique permitting the computation of multiple descriptors of shape and texture [18]. The relevant information present in the image is extracted using three classes of features, namely (i) shape, (ii) first-order, and (iii) texture-based features. First-order features are histogram-based and relate to the distribution of the grey-level values in the tissue. Shape features capture the geometrical properties of the region of interest (ROI), including volume, diameter, minor/major axis, or sphericity. Texture features are derived from images using five matrices that encode the global texture information. They aim to describe patterns using mathematical formulae based on the spatial arrangement of pixels.

Lately, CMR radiomics features have been utilised to appreciate the heart’s complexity derived from the left and right ventricles, revealing patterns invisible to the naked eye [14]. There are as yet no existing reports of clinical models based on PAT radiomics features, likely due to the absence of appropriate datasets.

PAT radiomics feature extraction

We used the PAT segmentation defined on the long-axis four-chamber images in the ED phase using the automated pipeline described above to derive our regions of interest (ROI) for radiomics analysis. We converted the contour points into binary masks, using a tool developed in-house, which we have made publicly available [19]. This software transformed delineated contour points for each ROI into a filled polygon in the coordinate space to form the binary mask. The harmonisation of the images was conducted using a histogram-matching technique applied to a reference image. The grey value discretisation was performed using a bin width of 25 to pull the intensity-based and texture radiomics features. The reference image for histogram matching was randomly selected, with careful consideration to ensure the chosen image did not contain artifacts. Histogram matching has been utilised effectively in prior radiomics-based models to standardise the intensity scale, thereby enhancing the model’s generalisation and classification capabilities [20,21,22]. In this study, imaging data was acquired using the same protocol (i.e., identical scanners and parameters) [17]. Therefore, we selected a single participant at random as the template for histogram harmonisation. This approach has been previously demonstrated to yield successful results in similar studies [20, 22]. The PyRadiomics platform (version 2.2.0) was adopted to extract 104 shape (n = 12), first-order (n = 17), and texture (n = 75) features from all PAT ROIs.

Feature selection

To mitigate the risk of multicollinearity and increase the interpretability of our models, after feature extraction we performed a correlation analysis among radiomics features. Pairs of features exhibiting a correlation coefficient with an absolute value of 0.8 or above were identified. From each pair, we removed one feature to maintain the distinctiveness of the predictors in our model. Following this correlation-based feature selection process, we retained 28 features from the original 104. We also included age and sex in our model as they are known to significantly influence cardiac health. For comparative purposes, we also developed another model which incorporated overall PAT area, age, and sex as predictors.

Predictive models

All the methods were implemented using Python version 3.9 and Scikit-learn [23] version 1.0.2. PAT radiomics features were used as predictors to classify prevalent HF and predict incident HF from non-HF controls. The features were normalised to zero mean and unit variance. We used seven binary classifiers followed by a voting classifier. We included the following classifiers to consider a wide variety of potential approaches: logistic regression (LR) [24], support vector classifier (SVC) [25], random forest (RF) [26], K-nearest neighbours (KNN) [27], decision tree (DT) [28], light gradient boosting machine (LGBM) [29], and multi-layer perceptron (MLP) [30]. To obtain each classifier’s optimal parameters, we used hyperparameter tuning and tenfold nested cross-validation, which consists of two loops [39,40,41] and different cardiomyopathies [42,43,44]. Pericardial fat radiomics represents an additional layer of information we can derive from standard of care CMR scans. Critically, our results demonstrate that radiomics can be used to discriminate HF cases from controls, signifying a potential novel avenue for better diagnostic and prognostic assessment.

Limitations

Our study provides initial insights into PAT radiomics for predicting HF though it has some important limitations. The models are preliminary and need further independent external validation. We used random undersampling due to extreme dataset imbalance, potentially leading to bias and overoptimistic performance estimates. While our models used a range of radiomics features and demographic variables, other relevant factors were not included, warranting comprehensive patient information in future research to enhance model performance.

Conclusions

Machine learning classifiers built upon radiomics features depicting the amount (larger PAT diameters) and texture character (greater tissue heterogeneity) of pericardial fat can be used to discriminate individuals with prevalent heart failure and predict incidence of future heart failure.