PCA-based polling strategy in machine learning framework for coronary artery disease risk assessment in intravascular ultrasound: A link between carotid and coronary grayscale plaque morphology

Highlights

• Coronary artery disease risk assessment in intravascular ultrasound.

• A link between carotid and coronary grayscale plaque morphology.

• Principal component analysis (PCA) for dominant feature selection.

• Classification accuracy of 98.43% and reliability index of 97.32%.

Abstract

Background and objective

Percutaneous coronary interventional procedures need advance planning prior to stenting or an endarterectomy. Cardiologists use intravascular ultrasound (IVUS) for screening, risk assessment and stratification of coronary artery disease (CAD). We hypothesize that plaque components are vulnerable to rupture due to plaque progression. Currently, there are no standard grayscale IVUS tools for risk assessment of plaque rupture. This paper presents a novel strategy for risk stratification based on plaque morphology embedded with principal component analysis (PCA) for plaque feature dimensionality reduction and dominant feature selection technique. The risk assessment utilizes 56 grayscale coronary features in a machine learning framework while linking information from carotid and coronary plaque burdens due to their common genetic makeup.

Method

This system consists of a machine learning paradigm which uses a support vector machine (SVM) combined with PCA for optimal and dominant coronary artery morphological feature extraction. Carotid artery proven intima-media thickness (cIMT) biomarker is adapted as a gold standard during the training phase of the machine learning system. For the performance evaluation, K-fold cross validation protocol is adapted with 20 trials per fold. For choosing the dominant features out of the 56 grayscale features, a polling strategy of PCA is adapted where the original value of the features is unaltered. Different protocols are designed for establishing the stability and reliability criteria of the coronary risk assessment system (cRAS).

Results

Using the PCA-based machine learning paradigm and cross-validation protocol, a classification accuracy of 98.43% (AUC 0.98) with K = 10 folds using an SVM radial basis function (RBF) kernel was achieved. A reliability index of 97.32% and machine learning stability criteria of 5% were met for the cRAS.

Conclusions

This is the first Computer aided design (CADx) system of its kind that is able to demonstrate the ability of coronary risk assessment and stratification while demonstrating a successful design of the machine learning system based on our assumptions.

Introduction

The major cause of morbidity in the world is due to cardiovascular disease (CVD). In 2012 alone, CVDs caused 17.5 million deaths worldwide, out of which, 7.4 million deaths were due to coronary arterial disease and 6.7 million were due to stroke or cerebrovascular disease [1]. A higher occurrence of CVD in the young and middle-aged population is observed in the south-east Asia region. About 35% of all such deaths are between the age group of 35–64 years and are estimated to happen in India [2] between the years of 2000 and 2030.

CVD includes coronary artery disease and cerebrovascular disease. These diseases occur due to atherosclerosis – a progressive and slow process of narrowing the artery, interrupting the flow of blood from the heart or to the brain. In severe cases, plaque deposits inside a vessel of the coronary artery and later ruptures causing myocardial infarction Fig. 1a (left).

The current-state-of-art methods for screening the severity of this disease is: computed tomography (CT), ultrasound (US), and magnetic resonance imaging (MRI). Due to radiation, CT may compromise the patients’ safety, but it is often used because it computes a calcium score in the coronary artery. Even though MRI was earlier not suited to show benefits for soft tissue characterization [3], [4], but now has started to be beneficial, but still lacks the concept of real time scanning. On the other hand, IVUS, though invasive, provides real time data, is less time consuming, and is less expensive [5], [6]. Though, IVUS is preferred over CT, because of CT radiation risk, both screening tools lack the ability to stratify risk based on plaque characteristics. This paper utilizes the novel idea of coronary artery risk stratification and assessment using the concept of the genetic makeup of the plaque in the coronary and carotid arteries (Fig. 1a (left, right)).

Honda et al. [7] and de Graaf et al. [8] recently pointed out the need for patients’ risk of severity prior to interventional procedures. This severity risk was linked to plaque morphology. Several authors have proposed that plaque in the coronary artery consists of several components such as: fibrous, fibrolipidic, calcified and calcified-necrotic, using modalities like CT, OCT, IVUS and elastography [9], [10], [11], [12], [13], [14], [15]. We thus hypothesize that different components offer different risk factors and when combined as a whole using the grayscale wall region image can be used for tissue characterization. We can thus leverage to study the morphological characteristics of these lesions and adapt a machine learning paradigm to predict the risk of severity of CAD leading to myocardial infarction. This paper explores the novel concept of morphological characteristics utilizing the coronary vessel wall region that possesses these plaque components.

It has been recently shown by many researchers that there exists a relationship between plaque burdens in the carotid and coronary arteries. Here we will discuss some of the key studies which correlate cIMT with risk stratification in cardiovascular events. It has also been a biomarker for cerebrovascular events (CVEs) [16], [17], [18], [19]. The relationships between coronary artery disease, cIMT and myocardial infarction have been demonstrated in previous studies. Ziembicka et al. [20] showed that there is a 94% chance of having coronary artery disease when cIMT > 1.15 mm. Ogata et al. [21] showed that maximum cIMT was highly correlated to left main coronary artery disease. Recently, Elias-Smale et al. [22] showed that cIMT > 1.26 mm can lead to myocardial infarction. Kao et al. [23] had shown that the cIMT > 0.80 mm (p < 0.01) lead to cardiovascular events. Our team lead by Ikeda et al. [24] showed that cIMT > 0.9 mm had a significant higher SYNTAX score, a risk indicator for coronary artery disease. The association between cIMT and coronary calcium volume has been recently shown by Suri's team [25], [26]. Another study by the same group has also shown the correlation between automated cIMT that includes bulb plaque and SYNTAX score was found to be 0.467 (p < 0.0001), compared to 0.391 (p < 0.0001) between sonographer's cIMT reading and SYNTAX score [27]. Thus, there is a clear relationship between cIMT and coronary artery disease severity. Based on the above analysis, we hypothesize that cIMT can be adapted for developing a link between the coronary plaque burden leading to coronary artery disease or carotid artery disease leading to stroke. Ahead, we will show how to use the cIMT from carotid artery (as ground truth labels) along with grayscale morphology of coronary artery for CAD risk assessment.

Ultrasound ability for real time tissue characterization has recently surfaced for stroke application by Suri's team [28], [29], [30], [31], [32]. This stroke risk assessment tool (AtheroRisk™, AtheroPoint™, Roseville, CA, USA) was adapted for the risk of vulnerability to plaque rupture. The AtheroRisk™ tool adapted a machine learning paradigm, which consisted of an offline (training-phase) and online (testing-phase) systems. The morphology-based tissue characterization was performed either on the grayscale cut-section images representing the plaque or the intima-media thickness wall region representing the total plaque area. The training-phase requires the ground truth and this consisted of binary label such as: either asymptomatic or symptomatic or a binary label which can be derived from the cIMT information measured by the sonographer. The training-phase used grayscale features and the ground truth labels to yield the offline training coefficients which were then used to transform the online grayscale features test set to predict the new risk labels. Such learning methods adapt a gold standard whose a priori information was known, such as which plaques were at high risk or low risk. A similar concept of characterization of plaque for classification of plaques into symptomatic and asymptomatic was developed by Suri's team (under the class of Atheromatic™ systems (AtheroPoint™, Roseville, CA, USA). This spirit is being extended to coronary artery disease application in this study for tissue characterization of the coronary artery wall region.

Our system uses coronary grayscale morphology for risk prediction of CAD severity using a machine learning paradigm. Since the numbers of grayscale coronary morphologic features are large, we use dominant feature selection using principal component analysis (PCA) using polling-based method. SVM is adapted for training and testing the machine learning algorithm. The stroke biomarker cIMT is adapted as ground truth for training the machine learning system which is then used for risk prediction of CAD. The cIMT biomarker threshold of 0.9 mm is used as a risk label for high/low risk CAD. A cross-validation approach is used for evaluating the efficiency of the machine learning infrastructure. Experiments are performed to study the effect of the data size on the classification accuracy and the role of PCA-based cutoff for feature selection during risk prediction. The overall system is novel and being used for the first time for risk prediction in CAD using PCA-based methods.

The layout of the paper is as follows: Section 2 presents the demographics and data acquisition. Section 3 describes the novel methodology of the proposed system for tissue characterization and risk stratification. Two experimental protocols are presented in section 4. The results are presented in section 5 that shows the optimization of SVM kernel type, dominant feature selection with change in PCA cutoffs and effect of data size on machine learning accuracy. The reliability, stability and feature retaining power are presented in performance evaluation section 6. A discussion on the results of proposed system is presented in section 7 and finally we conclude in section 8.