Prediction of acute coronary syndrome within 3 years using radiomics signature of pericoronary adipose tissue based on coronary computed tomography angiography

Acute coronary syndrome (ACS) can be often the first manifestation of coronary artery disease (CAD) and the main cause of death in the majority of the world’s population [1, 2]. As a widely used non-invasive imaging modality, coronary computed tomography angiography (CCTA) has shown its clinical value by enabling robust coronary plaque characterization and quantification [3, 4], especially for the identification of adverse plaque characteristics (APC) [5]. As plaque rupture is a complicated biomechanical process, whether the clinical outcome of vulnerable plaques developed into ACS may be affected by several factors. Among them, vascular inflammation is recognized as a key factor to both plaque formation and rupture, resulting in the occurrence of subsequent ACS [6]. A recent review of randomized controlled trials suggested that reduction of inflammation with colchicine on patients with CAD could lower the incidence of ACS [7]. Therefore, a comprehensive evaluation combining plaque characteristics with vascular inflammation may further enhance the prediction of ACS.

It has been established that there is a constant bidirectional manner between the vascular wall and pericoronary adipose tissue (PCAT); pro-inflammatory factors released from the inflamed vascular wall spread to PCAT in a paracrine manner to inhibit preadipocytes differentiation and lipid accumulation [8‐10]. Antonopoulos AS et al proposed a novel imaging biomarker, fat attenuation index (FAI), that captured the CT attenuation changes of PCAT and further revealed changes in PCAT composition induced by vascular inflammation [10]. Recent clinical studies indicated that PCAT attenuation measured from CCTA may contribute to identify high-risk plaque (HRP) progression and improve prediction for adverse cardiac events [11, 12]. Nevertheless, the changes in PCAT composition were not only related to vascular inflammation but also to dysfunctional adipose tissue remodeling characterized by fibrosis and vascularity [13, 14]. Thus, this approach only relying on CT attenuation to reflect changes in PCAT composition without considering complicated spatial relationship among voxels might lead to certain overlaps between pathologies.

Radiomics refers to the process of converting digital medical images into mineable higher dimensional data, the high-throughput extraction of quantitative image features for providing clinical-decision support [15, 16]. Radiomics was originally applied in oncology. Notably, it has been gradually applied to the coronary lesions and myocardium [17‐21]. Oikonomou EK et al [22] primarily proposed that a novel radiotranscriptomic signature of PCAT could detect additional disease-related changes in PCAT composition; meanwhile, their machine learning–powered radiomics analysis of PCAT could lead to a significant improvement of cardiac risk prediction. A recent report [23] based on CCTA-based radiomics characterization of PCAT surrounding coronary plaques found that there was a distinct PCAT radiomics phenotype between patients with acute MI and patients with stable or no CAD, yet the predictive value of PCAT radiomics surrounding plaques for future adverse cardiac events has not been mentioned. Therefore, our study aimed to develop a CCTA-based radiomics signature of PCAT surrounding coronary lesions to identify patients with future ACS within 3 years.

In this study, we developed an integrated score that incorporated radiomics features of PCAT surrounding target lesions and significant plaque predictors based on CCTA and validated the performance with respect to discrimination, calibration, and clinical application. This integrated score displayed superior diagnostic performance in the prediction of future ACS within 3 years compared with plaque score.

As many ACS events usually occur in patients without obstructive plaques, we should pay more attention to the identification of vulnerable plaques instead of the degree of luminal stenosis. The current CCTA could not only identify obstructive atherosclerotic plaques and plaque burden but also evaluate HRP features beneficial to risk stratification [4, 28]. Although the detection of HRP features has provided the incremental predictive value of coronary events to a certain extent [29], they are not a direct indicator of inflammation, just anatomical signs of vulnerable plaques and markers of the risk of rupture [4]. Recent studies [10] have suggested that coronary inflammation drives dynamic changes in perivascular adipose tissue (PVAT) composition, captured by a novel CCTA-derived imaging biomarker, the perivascular FAI, which reflects inflammation-induced PVAT changes in adipocyte size and lipid content are related to CT attenuation gradients. However, this metric based on CT attenuation simply reveals the average voxel intensity values without considering the complex spatial relationship among voxels. Radiomics enables help to provide the spatial distribution of voxel gray-level intensities and a quantification of heterogeneity [16].

In our study, we selected 14 most significant predictors from 107 radiomics features of PCAT based on CCTA using multivariate logistic regression analysis and found that patients of which PCAT with lower uniformity and higher heterogeneity were correlated with high possibility of future ACS, which indicated that heterogeneity of PCAT revealed by radiomics may reflect early pathophysiological changes in the adipose tissue around plaques; besides, the difference of PCAT radiomics parameters between two groups may be influenced by the related local inflammatory response, and PCAT radiomics characterization may be closely associated with subsequent plaque rupture. We further developed and validated that integrated score combining radiomics features with significant plaque predictors yielded a good diagnostic performance in identifying patients with ACS in both training and test datasets. Indeed, a recent prospective case-control study [23] suggested that a distinct radiomics phenotype of PCAT exists between patients with acute MI and patients with stable or no CAD, especially textural and geometric features that provided more additional information than the average attenuation of PCAT; moreover, their radiomics-based model outperformed PCAT attenuation-based model in identifying patients with MI. Oikonomou EK et al [22] developed a novel fat radiomics profile (FRP) using CCTA-based radiomics phenotyping of coronary PVAT and validated its performance in three different patient cohorts, which show that FRP contributed to further improve cardiac risk prediction for major adverse cardiac events beyond common established risk factors. Nevertheless, our study showed that the selected plaque predictors could not significantly improve the discriminating ability of integrated score in identifying patients with subsequent ACS compared with radiomics score, which could not mean that we did not need conventional plaque characteristics anymore to determine future ACS risk in the light of small sample size.

Based on our results, there was a statistical difference in MLD, MLA, DS, and HRP of plaque characteristics between two groups in the training dataset, but none of them was significantly different in the test dataset, which may be affected by the small sample size. Simultaneously, HRP was a more significant predictor among them. The present study [3] based on CCTA also showed that HRP was an independent predictor of ACS. Lee JM et al [30] also displayed that the culprit lesions had more frequent HRP than nonculprit lesions, and HRP further improved discriminatory ability in the identification of subsequent ACS. Nonetheless, our plaque score that was derived from MLD and HRP did not show a much better performance on predicting ACS outcomes in the test dataset. This might be due to the occurrence of ACS events is affected by many factors. Although coronary CTA has the ability to identify HRP features, there are still certain limitations in detecting small but vulnerable plaques which are potential to either rupture or rapidly progress to obstructive heart disease [31]. Moreover, HRP usually has a heterogeneous natural history and only a small proportion of them actually cause ACS events [32], several pieces of evidence suggested that half of ACS events happened without anatomically significant atherosclerotic plaques. Motoyama S et al showed that 83.7% of HRP identified by CCTA did not cause any ACS events [3].

Application of CCTA-based radiomics analysis to PCAT surrounding lesions in patients with ACS may increase our understanding of the related inflammatory response in the pericoronary environment. We developed a more comprehensive ML model for identifying patients with future ACS at a noninvasive imaging level by integrating radiomics features with plaque predictors, then found that the integrated score obviously added incremental discriminatory value in identifying patients with subsequent ACS over plaque score. The better performance of the radiomics-based score demonstrated that radiomics methods could extract more predictive information from PCAT based on CCTA than conventional plaque characteristics and have the potential to enhance the predictive ability of subsequent ACS. These findings further proved the important role of radiomics information in PCAT for the prediction of ACS.

There are still several limitations in our study. Firstly, we presented a retrospective case-control study design within a single center. The sample size of our work is relatively small; there is still a need for further external validation in an independent cohort to verify our findings. Secondly, although our work performed stratified analysis on three predictive models that indicated a good reproducibility and robustness exist, the application of our predictive models to general populations is limited to a single-center study. Several studies [33‐35] indicated that image acquisition, reconstruction, and analysis have a certain impact on the reproducibility of radiomics features. Thirdly, our study merely concentrated on PCAT radiomics phenotyping at a per-patient level, further study extended to a larger population at per-lesion level would be carried out. Furthermore, the current method of manually delineating ROI in our study is relatively time-consuming and easy to be affected by human factors, especially the process of image reconstruction greatly affected by the experience of radiologists. Although we did ICC to verify the stability of the extracted features, eliminate unstable features and try to build a robust model, we still hope to use semi-automatic or fully automatic segmentation technology to perfect this part of the work.

In conclusion, CCTA-based radiomics signature of PCAT could have the potential to improve the predictive ability of subsequent ACS. Radiomics-based integrated score significantly outperformed plaque score in identifying future ACS within 3 years.