What if a Sinus Rhythm ECG could show Atrial Fibrillation?
Artificial Intelligence has again penetrated the boundaries of medicine again, as new insights have revealed an AI-powered model that has shown adeptness at identifying patients who have intermittent (irregular) atrial fibrillation, even when the patient is in sinus rhythm whilst only using a 10 second test that is non-invasive, whereas intermittent atrial fibrillation formerly required weeks (if not longer) to be correctly diagnosed in patients. Albeit the study is in early stages, as is artificial intelligence in clinical healthcare, with further research, these models could be the difference in helping doctors to find the causes of enigmatic heart attacks and strokes, and ultimately guide them in giving correct treatments for patients who don’t fit into the traditional criteria.
The Mayo Clinic study, published by The Lancet, involved over 180,000 patients and is the first to deploy Deep Learning (a subset of Machine Learning) to identify patients with formerly overseen atrial fibrillation and showed an accuracy near 85%. The researchers’ aim was to train an artificial intelligence model to detect the foundations of atrial fibrillation in electrocardiograms taken from patients in sinus rhythm and that only lasted for 10 seconds. The electrocardiogram, albeit an incredible tool in Cardiology and found in similar abundances to paperwork in an office in a cardiology ward, is flawed in a notable manner. The human eye. In fact, the ‘invisible’ signals and abnormalities in the ECG that can be seen by the technology is often undetected by the human eye, but could lead to very important information about any potential presence of atrial fibrillation.
According to the Stroke Association, around 1.2 million people in the UK have atrial fibrillation. This figure rises up to at least 2.7 million in the US! Atrial fibrillation itself is actually simply a heart condition that results in an irregular (and generally unusually fast) heart rate, even pushing beyond 100 beats per minute at times which is well into the boundaries of tachycardia. In atrial fibrillation, the heart’s atria contact randomly and depolarise so quickly that the heart muscle cannot repolarise properly between contractions, which severely reduces the heart’s cardiac performance, output and efficency. This is due to the sino-atrial node’s natural rhythm being overridden by many different impulses rapidly firing at once in the atria, causing a fast and chaotic rhythm. It is as such associated with an increased potential of heart failure, stroke and even death, and often occurs along other heart conditions such as hypertension, atherosclerosis, heart valve disease and pericarditis and other non-cardiac diseases such as chronic obstructive pulmonary disease and pulmonary embolism. Since atrial fibrillation is not generally permanent, your heart can go in and out of this fibrillation, and it can often go un-diagnosed when using a single ECG as the heart may have been restored to sinus rhythm (normal). This obviously varies with chance (i.e the state of the heart when the ECG was taken) and also the type of atrial fibrillation, such as
paroxysmal atrial fibrillation – episodes come and go, and self-resolve in less than 2 days
persistent atrial fibrillation – longer episodes that span over a week unless treated
long-standing persistent atrial fibrillation – where atrial fibrillation has been continuous for over a year
permanent atrial fibrillation – where atrial fibrillation is continuously present without a limit
After enigmatic events as mentioned formerly, such as a stroke, it is vital to detect atrial fibrillation so corresponding patients are given anti-coagulation medicine such as Warfarin to reduce the risk of another stroke recurring, and other patients who are not suitable to anti-coagulation exempted. As modern healthcare has shown, it is important that patient-specific treatment is given to get the best possible results as your genetic conformation can be a potent factor. As we see in the field of Pharmacogenics, personalised medicine is the way that medicine is going as it is simply irrational to assume all humans share the same exact anatomy and biochemistry. Similar, but not exact, and this causes different factors to consider for each person. For example, the HER2 gene mutation causes about 30% of all breast tumours, by analysing these tumours, doctors can prescribe medications such as trastuzumab (Herceptin) and lapatinib which are drugs that shut down the activity of the HER2 gene, and could cut down deaths from HER2 breast cancer by up to 50%. Although not the same, but we can see how important it is to identify cause to give the correct response, and hence this is why identifying atrial fibrillation in an ECG is so important. Currently, identifying atrial fibrillation can be a timely event, at least weeks if not months needed with an implanted device (or regular check-ups, less rare) but this time factor could leave patients without treatment for extended periods of time, even if they are prone to a recurrent stroke that could be even more disastrous than the first…
However, as for the study, the coordinating researchers aimed to develop a neural network (remember that we are talking about Deep Learning here) to recognise subtle differences in a standard ECG that are potentially indicators of atrial fibrillation, that would not be noted by the naked eye. The neural network, which is a class of the deep learning architecture, scales with data quantity. Effectively, the more data you feed it, the better it performs, as we saw previously in Ophthalmology at the Moorfields NHS Trust with OCT scan interpretation by a very similar system. The ECGs of over 180,000 patients, equating to roughly 650,000 ECG scans taken in the last (roughly) 3 decades were stock piled into the system, ultimately sectioning the data into patients who did or did not have atrial fibrillation. This data was allotted into three training, internal validation and testing data sets for different purposes and measures of the system’s ability. It was clear that the system was very accurate indeed, by analysing only a single 10-second ECG in Sinus Rhythm, it was accurate to 79% and this value rose to 83% when multiple scans of the same patient were used, which is an incredible value for a system in such early development. Now researchers hope to carry out specific research on populations, such as patients with an unexplained stroke, and seeing how the system could genuinely fill in gaps in cardiology, as we could unearth new ways to explain strokes and myocardial infarctions that would otherwise stump cardiologists in a clinical environment today. For example, ESUS (embolic stroke of undetermined source) can be an immensely time and resource consuming process to figure out the puzzle of unexplained cardiac events, but atrial fibrillation could be the detectable factor that may provide answers , if not at least a plan of treatment (such as anti-coagulants for stroke victims with atrial fibrillation).
This application of AI in healthcare has been bolstered by the support of influential institutions such as Mayo Clinic. In fact, Dr Paul Friendman, Chair of the Department of Cardiovascular Medicine at Mayo Clinic had the following to say; "Applying an AI model to the ECG permits detection of atrial fibrillation even if not present at the time the ECG is recorded. It is like looking at the ocean now and being able to tell that there were big waves yesterday. Currently, the AI has been trained using ECGs in people who needed clinical investigations, but not people with unexplained stroke nor the overall population, and so we are not yet sure how it would perform at diagnosing these groups. However, the ability to test quickly and inexpensively with a non-invasive and widely available test might one day help identify undiagnosed atrial fibrillation and guide important treatment, preventing stroke and other serious illness."
The coordinators of the study hope that one day, their study could set the basis for other technologies as a point-of-care/immediate diagnostic screening tool to aid a doctor in assessing complex high-risk minorities. Screening people with hyperglycemia and hypertension for AF could aid in preventing health issues that result from the time and effort it takes for current screening methods, and not to mention the cost and complexity of the case, i.e a win-win situation.
The creators note a few confinements and accept the need for some further research and time before their systems are clinically introduced. The population studied, which were ECGs from cardiac patients, will have had a higher prevalence of atrial fibrillation as opposed to the general populated and so the system may need to be reconfiguration before a nationwide system can be introduced to assess a broader general population as the AI has been trained to reflectively classify clinically demonstrated ECGs, and may not yet be calibrated for use with healthy patients or those with unexplained cardiac notables, such as stroke. Ultimately, the AI is dependent on the quality of the data sets it was provided, and this could be of concern given the (near) three decade period that they were taken from. Even if the AI can diagnose what the doctor’s eye missed, there is always a risk (as with any healthcare analysis) of being wrong, and further testing and research are definitely needed before we can safely introduced these technologies into our hospitals for patients, on one hand, delaying the introduction of life-changing technology but on the other hand, we might be implementing an unstable system into the healthcare of our most vulnerable patients. Only time will tell how these systems sway, but early results are incredibly admirable.
To end the article, I will leave you with two quotes from two key figures;
Dr Xiaoxi Yao, a study investigator said "It is possible that our algorithm could be used on low-cost, widely available technologies, including smartphones, however, this will require more research before widespread application."
Dr Jeroen Hendriks, of University of Adelaide and its Hospital (Royal Adelaide Hospital) said: "In summary, Attia and colleagues are to be congratulated for their innovative approach and the thorough development and local validation of the AI-enabled ECG. Given that AI algorithms have recently reached cardiologist level in diagnostic performance this AI-ECG interpretation is ground-breaking in creating an algorithm to reveal the likelihood of atrial fibrillation ECGs showing sinus rhythm."
As always, a great thanks to these fantastic institutions and even more fantastic persons. To Dr Paul Friendman, Dr Jeroen Hendriks, Dr Xiaoxi Yao, the Mayo Clinic, and the University of Adelaide, thank you for your fantastic efforts and I wish you the very best in your future research. You are taking medicine into the next phase, where man and machine co-exist to new heights to truly boast the wonders of AI-assisted clinical healthcare.