By Rashed Karim

Research Fellow at King’s College London School of Biomedical Engineering and Imaging Sciences

It’s exciting to envisage that future treatments for cardiovascular disease will be supported by intelligent systems and devices. At the NIHR Biomedical Research Centre at Guy’s and St Thomas’  and King’s College London we are constantly looking at using the latest technologies to enhance and support quicker, more efficient diagnosis and more accurate treatments for patients with cardiovascular disease.

The burden of cardiovascular disease on the NHS and patients is significant. In the UK alone, 7 million people are affected by cardiovascular disease at any one time. Most common conditions include heart attack, coronary heart disease, heart failure and atrial fibrillation. The total annual healthcare cost of cardiovascular disease is £9 billion a year, with an estimated disease cost of £19 billion to the UK’s economy.

We think that using machine learning in the field of imaging could be a game-changer for these conditions. Currently imaging is a vital tool in diagnosing and treating heart disease – since the heart is buried deep in a patient’s chest, imaging helps clinicians to understand a patient’s condition and plan their treatment. But analysing images takes up vast amounts of clinicians’ time. If we could use machine learning to do some of this work, this could allow us to start patients’ treatment sooner and therefore improve success rates. Another benefit would be freeing up time that clinicians currently spend analysing scans and therefore reducing costs for the NHS.

Can we teach machines to read hearts?

We have focused our research on using machine learning to diagnose and treat Atrial Fibrillation (AFib). AFib, when the heartbeat is irregular and often abnormally fast, is one of the most common forms of abnormal heart rhythm and a major cause of stroke. There are 1.3 million people diagnosed with the condition in the UK, with a further 500,000 estimated to be living with undiagnosed AFib.

Most patients with symptoms of AFib are diagnosed via Magnetic Resonance Imaging (MRI) – a 3D scan of the heart. These images are complex, showing the muscle tissue of the heart, its chambers and the special muscle bundle network that tell the heart when to beat. To create these 3D pictures, hundreds of images are taken and put together. Because of this, the volume of data acquired during these scans is vast, in the region of gigabytes. Processing all of this data in a timely manner has become a real challenge.

Usually the first step in analysing a 3D cardiac MRI scan is segmenting and classifying important information within an image. In AFib, the left atrium, which is a complex upper chamber of the heart, needs to be identified and outlined by a radiologist before it can be fully appreciated in 3D. This can take a very long time and sometimes a second type of scan is necessary to get enough information to plan treatment. If the atrium can be outlined from just one scan, clinicians can put together a treatment plan sooner, deliver it to patients quicker, and with fewer hospital visits.

We have been working to develop more efficient ways to read these scans, using machine learning and application of computing systems known as artificial neural networks (ANNs). ANNs are inspired by the biological neural networks that make up the brain. This means they could ‘learn’ how to read the images just like an expert radiologist. But they could read these images far more quickly and around the clock. Humans need to sleep and eat occasionally, but machines don’t.

We have used a special type of an ANN which can learn the anatomy of cardiac chambers from hundreds of MRI scans that have been manually annotated by consultant clinicians. The entire network can be trained on thousands of manually annotated images in less than half a day. Once trained, it can help us segment an MRI image in less than a minute.

Using these imaging techniques and machine learning will assist clinicians in detecting cardiac conditions with greater accuracy over the currently applied diagnostic tools and approaches, and in a fraction of the time. The more accurately and quickly we can understand patients’ conditions, the more likely it is that our treatment will be successful.

Guiding treatment with images

After they have been diagnosed with heart failure, part of treatment for many patients is to undergo cardiac resynchronization therapy (CRT), where a cardiac pacemaker is implanted. Pacemakers are small, credit-card sized electrical devices that are implanted in patients’ hearts to regulate their heart beat if it is too fast, too slow, or irregular. CRT pacemakers have three electrode leads which are implanted into the heart tissue – they sense the heart’s beat, and send electrical signals to stimulate a regular heartbeat.

25,000 people a year in the UK have a pacemaker fitted to treat a variety of heart conditions, including heart failure. This number has increased over the last 30 years. Statistics show that with current diagnostic and treatment methods, 30-50% of patients do not benefit from pacemake implants, with many patients requiring a repeat procedure.

One key factor for success is placing the pacemaker and the leads in the optimal place, and the best site is specific to each patient. Most surgeons, however, place the lead in the same location for every patient.

Our hypothesis is that if we can find areas of the heart that will respond to the pacemaker poorly and ensure the electrodes are not implanted in these areas, we can increase the chance of the treatment being successful the first time, and reduce the number of patients having to have repeat procedures. For instance, one region we know has a poor response rate is scar tissue from previous heart attacks.

We aim to make the implantation of a pacemaker more tailored to each patient by using image technology to guide therapy.

Working in collaboration with Siemens, we have developed a technique to identify regions within the heart that respond best to signals from a pacemaker. These regions are automatically identified by a computer algorithm and classified into varying degrees of responsiveness. The computer algorithm analyses the MRI images and makes detailed measurements of various properties of the heart muscle tissue. Together these measurements can help clinicians identify the best place for a pacemaker in each individual patient.

The use of intelligent systems and technical innovations like artificial neural networks and image guided therapies have the potential to significantly improve outcomes for patients, while reducing the cost of cardiovascular disease for the NHS.