Artificial intelligence: a new clinical support tool for stress echocardiography

KEYWORDS:

• Echocardiography
• stress echocardiography
• artificial intelligence
• machine learning
• automated analysis

1. Introduction

Echocardiography remains the imaging modality of choice for the early detection and diagnosis of cardiovascular disease because it is portable, non-invasive, radiation-free, and allows real time imaging of the heart. Furthermore, echocardiography is relatively inexpensive when compared with other imaging modalities and so is accessible in the majority of healthcare settings around the world [1]. However, accurate diagnosis using echocardiography requires a high level of clinical skill and operator training to ensure good quality image acquisition, optimization and interpretation. Wide implementation of echocardiography guidelines have helped standardize these processes and ensured reproducible echocardiographic parameters. However interpretation remains dependent on operator experience and a limited set of echocardiography parameters [2]. Computational tools that allow complex, standardized analysis and quantification of images have emerged, which provide more comprehensive characterization of cardiac structure and function [3,4]. However, it is the combination of these approaches with artificial intelligence tools, such as deep learning, which can form the foundations of a new era of consistent and accurate echocardiography image interpretation.

2. What is artificial intelligence?

The first applications of artificial intelligence in healthcare were reported over three decades ago [5]. However it is only in the last few years, as artificial intelligence has become embedded within multiple areas of life, that there has been an exponential growth of interest in whether it can assist in automated diagnosis and personalized patient management. Artificial intelligence includes computational techniques that ‘learn’ from existing data to make future decisions. Deep learning is a method composed of many layers of highly interconnected processing elements, which are able to represent high levels of abstraction. The use of deep learning with imaging data is usually based on convolutional neural networks that mimic, to some extent, how the human ventral stream is structured [6]. These techniques facilitate rapid analysis of massive amounts of data [7]. Less fluid computational approaches are also possible such as support vector machines and random forests [8–10]. However, the objective of all these methods is to learn patterns from existing sets of clinical data, such as clinical notes, blood test results or images, to allow future sets of data to be automatically processed [5]. In medicine, applications of artificial intelligence have been innovative, particularly in medical imaging [11]. Medical images contain large sets of data that require intensive training and experience in order to detect abnormalities [6]. For clinical adoption, it is important the true impact of artificial intelligence systems, compared to operator-led analysis, on patient outcomes, including how changes in workflow and test accuracy impact on health economic costs, needs to be validated in clinical trials. However, machine-assisted interpretation of medical images offers the potential for more consistent decision-making that could improve patient outcome [11].

3. Artificial intelligence in echocardiography

The application of artificial intelligence in the clinical practice of echocardiography has been less advanced than in some other areas of medical imaging. Every echocardiogram generates multifaceted and complex information within the image, which is mostly filtered by the eye of the operator when being interpreted or measured. Therefore, potentially useful data that could be used for quantification of cardiac structure and function, or used for diagnosis, may be missed or overlooked [8]. Recent applications of artificial intelligence in echocardiography have shown promise in the field of automated image selection and quantification. The left ventricle appears in multiple echocardiographic views and a deep learning model was able to recognize 15 major transthoracic echocardiography views accurately, including continuous and pulsed wave Doppler traces [12]. Automated quantification or border recognition of left and right ventricular function could then be possible using demonstrated techniques [13,14]. Valvular morphological quantification also appears to be possible using automated machine learning analysis of 3D transoesophageal echocardiography images of the mitral valve. From this analysis it was feasible to achieve reproducible measurements of mitral valve annulus without significant user intervention [15]. Image interpretation is a distinct task that may also be tackled with artificial intelligence approaches. Machine learning models have provided efficient differentiation of cardiovascular hypertrophic phenotypes including those with hypertrophic cardiomyopathy and athletes [8]. Classification of constrictive pericarditis and restrictive cardiomyopathy has been shown to be possible particularly when conventional echocardiography parameters were combined with parameters obtained using speckle tracking echocardiography [10].

4. A new medical device for stress echocardiography?

Stress echocardiography is a specific test that studies how cardiac function changes after a patient is exposed to a stressor, either exercise or drugs such as dobutamine [16]. Stress echocardiography is the most widely used functional test for coronary artery disease with nearly four million performed in the United States every year [17]. Current interpretation is based on visual assessment of the images by an experienced operator. The test typically has a sensitivity and specificity of around 80% for identification of functionally significant coronary artery disease when assessed by angiography or clinical outcome [18]. Machine learning algorithms hold the potential to utilize the whole dataset from each echocardiography image, and detect abnormal myocardial segment patterns to ensure more accurate and consistent results. Thereby, these systems might help to reduce the number of inconclusive test results also leading to potential time, effort and cost savings for the patients, as well as the clinical care team. Attempts to apply machine learning techniques to stress echocardiography have shown promise. These included approaches to automate quantification of wall motion with machine learning classification [19] or to extract and categorize information from processed data such as principal strain maps [20]. These methods were able to achieve sensitivity or specificity comparable with human operators [18]. However, a limitation in improving the accuracy of these methods further and, ultimately, translating these techniques into a medical device has been a lack of sufficient high quality data for both training and validation. Over the last 7 years, we have been compiling image datasets from patients undergoing stress echocardiography linked with longitudinal clinical outcome data, as part of an ongoing clinical research study [16]. Advances in machine learning and, in particular, deep learning, during this time have meant it has become possible to use a multi-parametric assessment of cardiac function, based on 1000s of image-derived features at rest and stress, to predict patient outcome in longitudinal clinical datasets of sufficient quality and size for medical device development.

5. Conclusions

Echocardiography remains the main imaging modality in the diagnosis of cardiovascular diseases. The application of artificial intelligence into echocardiography offers the potential for accurate and reliable image identification, quantification and interpretation. Machine learning systems may reduce image analysis time, expedite clinical decision-making and provide iterative feedback to train less-experienced clinicians [11]. The first tangible clinical applications of artificial intelligence in echocardiography are now emerging, particularly in the field of quantification. Over the next year we expect the first applications of artificial intelligence in stress echocardiography to emerge as a new clinical decision support medical device.

Declaration of interest

The author has no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Additional information

Funding

P Leeson acknowledges support from the British Heart Foundation, Oxford BHF Centre for Research Excellence and National Institute for Health Research Oxford Biomedical Research Centre, and he is a shareholder of Ultromics Ltd. M Alsharqi acknowledges support for her studentship from the Saudi Arabian Cultural Bureau. A Mumith is an employee of Ultromics Ltd. R Upton is the CEO and a shareholder of Ultromics Ltd.