By Sophie Arthur
February 11, 2019
Time to read: 4 minutes
Current statistics from the British Heart Foundation indicate that heart and circulatory disease accounts for more than a quarter of all deaths in the UK. That equates to 160,000 deaths a year or one death every three minutes. One way we can try to treat and protect against life-threatening heart attacks or strokes is by trying to more accurately assess each patient’s individual risk.
Cardiac imaging forms an important part of the initial assessment of patients suspected of having heart failure. Currently, doctors evaluate a patient’s risk of heart disease by taking simple measurements of the volume and mass of the heart – but this approach is not accurate or patient-specific.
We live an increasingly digital and data-driven world, from our wearable devices to the wealth of information available from health records and genetics. But what role could Artificial Intelligence (AI) and all of this data play in healthcare?
A team at the MRC London Institute of Medical Sciences (LMS) have found a way to harness AI to enable doctors to predict outcomes for heart patients more accurately and find the best treatment for individual patients.
Research published in the journal Nature Machine Intelligence by Declan O’Regan, of the MRC London Institute of Medical Sciences (LMS), reports on a machine learning tool that can accurately predict a patient’s risk of heart failure by tracking the motion of their heart from cardiac MRI scans without needing any human involvement.
In the study the researchers used the technology to predict the prognosis for 302 people with a heart condition called pulmonary hypertension. Patients with pulmonary hypertension were chosen as the choice of treatment is dependent on the individual patient’s risk classification. The technology correctly predicts a patient’s prognosis 75% of the time and outperforms doctors’ measurements.
The AI in this study uses computer vision to analyse the motion of the heart in a sequence of MRI scans. The algorithm developed is trained on a series of patient MRI scans and tracks the motion of the heart at hundreds of points every second – creating an incredibly detailed map of how the heart moves. From these 3D pictures of the moving heart the machine then learns to predict the risk of dying from heart failure by finding simple patterns hidden in the data. The more images and data the algorithm is given the more it can learn about the intricate features of heart movement that are important. This strengthens its ability to spot the earliest indications of a future life-threatening cardiac event.
Image shows a series of loops protruding from a rotating model of the heart. Each loop shows the path that that particular part of the heart moves through when the heart contracts and relaxes. The colour of the loop indicates how quickly that movement occurs too, with red loops contracting more quickly than blue loops.
Declan O’Regan, senior author of this paper, discussed the ambitions for this technology:
“Our ultimate goal is to see this technology used throughout the NHS, not just for cardiac events but for other applications too. But first we will be evaluating the algorithm on larger cohorts of cardiac patients, in collaboration with centres in the UK and Europe, to see how well it performs in a real-world environment. This would just be using motion analysis from cardiac MR images, but there is so much more data out there that can enrich this technology. Incorporating a patient’s health records, genetic information, metabolic signature or even the heart data from your wearable device can give a much more precise and personalised recommendation of treatment.”
This exciting new imaging technology is not only the most precise prediction of future cardiac events yet, but crucially still allows doctors to interpret the outputs from the algorithm. This is the next step in allowing clinicians to tailor and guide a treatment option that is personalised to each patient.
‘Deep learning cardiac motion analysis for human survival prediction’ was published in Nature Machine Intelligence on 11 February 2019. You can read the research article in full here.