Using artificial intelligence to study atherosclerosis, predict risk and guide treatments in clinical practice

Abstract

Despite advancements in cardiovascular diagnostics and therapeutics over the last few decades, coronary heart disease is still the number one killer globally. We would all accept that typically, we detect coronary artery disease only if it is symptomatic. Angina is what attracts all the attention of the healthcare systems, leading to several tests [functional imaging tests for ischaemia, coronary computed tomography angiography (CCTA) scans, invasive angiograms, etc.] and eventually to secondary prevention measures (pharmacological and/or expensive revascularizations). What we usually ignore, as clinical cardiologists, is that coronary artery disease does not start with angina. The haemodynamically significant plaques that give the symptoms are usually calcified (hence not prone to rupture), and they are not the ones that put the patient at risk of heart attacks; indeed, only 50% of the heart attacks happen in patients with flow-limiting disease, and even in those cases, the ruptured plaques are usually not the ones that cause the obstruction.1 It was not surprising that the ISCHAEMIA trial2 found the targeting of flow-limiting plaques invasively non-superior to medical therapy in prolonging life, given that the later targets/stabilizes all plaques (significant or not). This introduces a series of questions for the cardiovascular community: (i) can we identify the obstructive plaques that are prone to rupture, and selectively treat them invasively to improve prognosis? (ii) can we identify the small but vulnerable plaques before they rupture (i.e. identifying the ‘vulnerable patient’) and stabilize them by pharmacological treatments? (iii) can we do even better by detecting the coronary arteries that are most likely to develop high risk atherosclerosis (i.e. identifying the vulnerable ‘healthy’ individuals before they develop disease), allowing pharmacological and other interventions to prevent disease from developing altogether? Although the causes of atherosclerosis are multifactorial, it is well-accepted that vascular inflammation is a major driver of atherogenesis. Vascular inflammation not only plays a major role in the formation of atherosclerotic plaque, but also triggers plaque rupture, hence it is considered a rational therapeutic target to reduce cardiovascular risk as shown in randomized trials with canakinumab (Cantos trial) or colchicine (LoDoCo2 and COLCOT trials). Therefore, to enhance cardiovascular disease (CVD) risk prediction and provide precision treatments to patients, it is essential to assess the residual inflammatory risk. CCTA is now the first-line investigation for stable chest pain,3,4 although >75% of the CCTAs performed worldwide do not present any haemodynamically significant disease and are archived without any action by the clinical care teams. More than 150 million computed tomography (CT) scans are performed globally each year and it is estimated that by 2025, 10%–15% of all CT scans will be CCTAs. In the UK, ∼250,000 CCTAs are performed yearly, and this number is expected to grow to 350,000 in the near future if the National Institute for Health and Care Excellence guidelines are fully implemented.5 The recent inclusion of CCTA as the first-line investigation for stable chest pain (Class I indication) in the ESC and the AHA/ACC guidelines4 is expected to have a similar impact on the growth of the numbers of CCTA scans performed globally. The growing number of these scans available, generates unique opportunities to detect and risk-stratify our patients based on the images of their coronary arteries. Artificial intelligence (AI) offers new opportunities to extract information from medical images, which are not visible to the human eye, while it can be used to perform faster and better, repetitive processes traditionally done by human operators (such as image segmentation, classification, etc.).5 The quantification of atherosclerosis burden from routine CCTA scans requires the segmentation of coronary plaques which is time-consuming and operator-dependent. In addition, the visual interpretation of the anatomical high-risk plaque features (i.e. positive remodelling, low-attenuation plaque, napkin ring sign, and spotty calcification) is prone to under- or over-reporting even by experienced imaging specialists. AI is currently applied on CCTA images to automate the process of coronary plaque detection, segmentation, classification, and quantification of its different structural components (i.e. calcified, lipid-rich, fibrous, thrombus). Plaque classification is achievable by using deep learning approaches as well as by applying more structured data extraction processes like texture radiomics. Indeed, with texture radiomics, we can extract thousands of features from a segmented image volume, which describe the ‘texture’ of the tissue under investigation, and machine learning methodologies use these features to characterize the tissue of interest (e.g. the segmented atherosclerotic plaque). Extensive characterization of coronary atheroma is now clinically available, through various AI-enhanced reporting solutions and services (Figure 1). However, the wealth of information on plaque characteristics leads to another big problem: how do we interpret the different quantitative and qualitative plaque characteristics in clinical practice? What would trigger change of management and when? In clinical practice we are used to treat risk, so it is important to be able to translate this information into absolute risk, and then follow the clinical guidelines to guide management. But is it enough to know the structural plaque characteristics, to provide reliable risk prediction? Structural plaque characteristics do not provide any indication of the inflammatory plaque burden that triggers plaque-rupture events. It is, therefore, essential to have clinical tools that would allow accurate detection of the inflamed atherosclerotic plaques from routine CCTA scans and take this into account in refining the patient’s absolute risk. Artificial intelligence-assisted image interpretation for the assessment of coronary artery disease, prediction of patient’s risk and deployment of personalized medical management. Computed tomography images derived from the CaRi-Heart® v2.5 medical device. It has recently been shown that perivascular adipose tissue (PVAT) changes its biology and structure as a result of vascular inflammation and this results in respective changes in PVAT phenotype (an increase in the water: lipid ratio, and reduction in adipocyte size close to an inflamed artery).6 Since adipose tissue and the water: lipid phases can be detected on CT, shifts in PVAT attenuation as a result of the underlying vascular inflammation can be quantified non-invasively by AI post-processing of routine CCTA images. An imaging biomarker of PVAT, the fat attenuation index (FAI) captures and tracks 3D gradients in the attenuation of PVAT around the coronary arteries and provides information on the presence of coronary inflammation from the post-processing of routine CCTA scans and the help of AI algorithms.6 Pericoronary FAI can reliably discriminate between stable and unstable plaques in acute coronary syndrome patients,6 and can be used to detect the patients at risk for a heart attack.7 A clinically used algorithm (FAI score) describes the overall inflammatory burden of a coronary artery, and is projected on age- and sex-adjusted nomograms for clinical interpretation.8 This information is integrated into a prognostic model together with clinical risk factors and metrics of atherosclerotic plaque extent, resulting in a powerful tool that predicts future cardiac events, even in patients with no evidence of structurally visible atherosclerosis at the time of the CCTA (Figure 1).8 This ability to predict risk for future cardiac events from CCTA images, is further enhanced by applying the radiomic analysis of PVAT.9 Building on the success of radiomics, ‘radio-transcriptomics’ provides a concrete link between radiomic disease markers and genes that are differentially expressed in diseased tissues. The transcriptomic profile of a tissue (derived from RNA sequencing) can form the ‘molecular ground truth’ against which radiomic models are trained by using machine learning approaches.5,9 The vascular inflammation-driven changes in PVAT (such as fibrosis, angiogenesis, lipolysis, and oedema) exhibit distinct textural features in CCTA, captured using such a radio-transcriptomic approach. Indeed, we have recently constructed a powerful radio-transcriptomic signature (C19RS) that reflects cytokine-driven arterial inflammation by analysing the respective perivascular radiomic changes.10 When C19RS was tested in patients with COVID-19 infection, it allowed accurate prediction of in-hospital mortality but most importantly it identified those patients who would respond well to anti-inflammatory treatments such as dexamethasone.10 In conclusion, the AI-assisted analysis of CT imaging data can revolutionize CVD prevention; the integration of patient’s risk factor profile, quantitative metrics of coronary plaques and the coronary inflammatory burden using AI, allow accurate prediction of future cardiovascular events, and can guide prevention strategies. Such a cardiovascular risk prediction tool as recently been developed for clinical use, and it is now being recalibrated in the ‘Oxford Risk Factors And Non Invasive Imaging Study’ (ORFAN Study), a study of 250,000 patients with CCTA scans linked with 15-year cardiovascular outcomes. This individualized AI score of coronary risk prediction could ultimately replace the existing prognostic models based solely on clinical risk factors (e.g. the ESC SCORE2), when CCTA information on atherosclerotic disease activity is available. The immediate future will see various AI-assisted solutions interpreting the images, integrating imaging information with multidimensional clinical and demographic data and guiding strategies for the prevention and treatment of CVDs. Personalized cardiovascular medicine is the ultimate price of AI in clinical practice. C.A. is supported by the British Heart Foundation (CH/F/21/90009, TG/19/2/34831 and RG/F/21/110040), Innovate UK (grant 104472 and the National Consortium of Intelligent Medical Imaging (NCIMI)) through the Industry Strategy Challenge Fund (Innovate UK Grant 104688). P.P. is supported by an Industrial Fellowship (Royal Commission for the Exhibition 1851) with the University of Oxford as the Academic partner and Caristo Diagnostics as the Industrial partner. A.S.A. is supported by the Hellenic Foundation for Research and Innovation (HFRI, grant Number 00468).

Topics & Concepts