Mr Andrew Fletcher, Cardiac Clinical Scientist at Royal Papworth Hospital NHS Foundation Trust, talks about the next wave of advances in artificial intelligence and what this could mean for healthcare scientists.
Imagine this… You start your transthoracic scan by acquiring a single 3D volumetric dataset of the whole heart with an ultra-high image resolution next-generation ultrasound machine. You inspect each valve with 3D colour and Doppler, perform some tissue-Doppler, then view the aorta, pulmonary artery and inferior vena cava. During this process, your saved images are analysed by cutting-edge software in the background. By the time you finish your abbreviated scan, a range of automatically made and accurate measurements are available for viewing on your machine, so that you can expand your image-set if required. Overall, your scan was quicker, you felt less neck/back/arm-ache, you had more time for interacting with your patient and your measurements are ready for reporting upon.
Think I’m talking science fiction? Well, advances in computer science and developments in artificial intelligence (AI) mean that this kind of workflow could be far closer than we think. The fast moving pandemic of the last 12-months often required us to focus clinically on the here-and-now, but this national Healthcare Science Week I encourage you to take a moment to glance into the future.
AI is often thought of as a computer doing something that normally requires human levels of intelligence whilst machine-learning (ML) is a subset of AI, where simply put the software learns to perform a task (see image 1). AI and ML are two of the foremost buzzwords in the current ‘digital revolution’ sweeping across all areas of medicine, but how exactly might they impact us as healthcare scientists in echocardiography?
It’s widely appreciated that advanced echocardiographic metrics like 3-dimensional volumes, ejection fraction (EF) and global longitudinal strain (GLS) have significant diagnostic and prognostic value, but historically some of the biggest challenges to their widespread adoption clinically have been the extra time required to perform analysis, and the lack of perceived accuracy (particularly when image quality is suboptimal, like for many of our patients!). But coupled with improvements in ultrasound technology where frame rates of up to 1000Hz can now be obtained with prototype machines, AI may be our new best friend in overcoming some of these longstanding adoption issues.
Already we are seeing ultrasound machines and reporting software with automated or semi-automated parameter measurement and calculation. However they don’t always work perfectly, requiring manual editing which introduces bias and variability. But if, as has been shown in research studies and is now available commercially, AI can perform the mundane measurements aspects accurately for you, on better quality pictures due to the ultra-high frame rate, then you could focus more time upon interpretation and placing your findings into the clinical context, right?
The upcoming wave of AI advances could offer so much more than this though. We currently use only a small proportion of the information that is contained within each scan. Measurements are taken at specific locations and at specific time points, for example a wall thickness or a tissue-Doppler e’ velocity. Hence, AI has a distinct advantage over us in that it can simultaneously process huge amounts of information, from all corners of the image, across the (potentially) hundreds or thousands of frames of the whole cardiac cycle. This rich treasure trove of information could contain useful new parameters which could vastly improve diseases diagnosis or predict patient deterioration.
AI could also help us manage the large amount of parameters we often seem to end up with. For example, integration of resting and exercise-echocardiographic parameters by ML has been reported to improve identification of heart failure (see Sanchez-Martinez et al. 2018, Circulation Cardiovascular Imaging). But AI wasn’t replacing them performing the resting and exercise scan, it wasn’t encouraging the patient to reach target heart rate like we do, it wasn’t acquiring those challenging images during heavy breathing of the patient. It can though support us with data processing and analysis, making better use of all the information we have worked hard to obtain, and reducing variability between different staff.
What I find particularly cool about AI is that it often excels in detecting pathology because it can pick up upon subtle patterns and hidden features in data. A particular favourite example of mine is from the Mayo Clinic in the USA (Attia et al. 2019, Nature Medicine) where a ML algorithm was trained with paired 12-lead electrocardiogram (ECG) and echocardiogram data to identify severe left ventricular systolic dysfunction. When presented with new ECGs, it was 86% accurate for identifying dysfunction. Furthermore, the potential for AI to act as a screening tool was demonstrated in their study, because patients with a positive ECG result (but normal echocardiogram) were at four times higher risk of developing dysfunction in the future than those with a negative ML result.
Far less talked about is where AI performs inferiorly or similarly to us mortals. I’ve seen a few examples but as always, the scientific community isn’t very good at publishing and publicising negative results. Rest assured, as cardiology healthcare professionals we are still pretty awesome at what we do, and there will always be a need for our input, oversight, experience and reasoning. So don’t worry about being replaced by Alexa or Siri just yet…
As a clinical scientist in echocardiography, I’m sure I’m not the only one who would like to realise some of the benefits outlined above. Ultrasound technology advancement and AI analysis could really help us, so I do hope you find the prospect of embracing them as exciting as I do!