Health Systems Action

Eagles, beagles and a moose – change comes to cardiology

This week I attended a webinar series hosted by the Cardiac Arrhythmia Society of Southern Africa (CASSA), taught by cardiologists from the University of Cape Town, focused on Essential ECGs. These are the ones you can’t miss – if you do, your patient might die!

I surprised myself by getting the quizzes mostly right. Surprised because, apart from periodic ACLS certification updates, the last time I formally studied ECGs was years ago, though a while since the 1903 invention of the electrocardiogram itself. The course is great, but gave me the feeling nothing much has changed in the way we read – or are taught to read – the tiny yet critical electrical signals of the heart’s electrical activity. Was I correct?

Willem Einthoven inventing the ECG while taking a foot bath. He got clean feet and a Nobel Prize.  Image: Wikipiedia

Mostly right isn’t good enough

In clinical practice, “mostly right” often isn’t good enough, especially in potentially life-threatening scenarios like the ones covered in the webinars.

But ECG interpretation is hard. A large (and still growing number) of criteria and rules defined by empirical associations between the ECG and disease conditions (such as heart attack and severe rhythm disturbance) are described for the practitioner to remember and use to identify disease.

A meta-analysis of 78 studies showed that with training, accuracy of ECG interpretation improves but it starts pretty low (54%) and gets only to 67%, across a range of medical professionals, 75% among cardiologists1.

Training from experts, at home, for free

The first change, and a welcome one, is that I can attend training for free and learn from the comfort of home thanks to internet technology we now take for granted, the talent of our academic colleagues and the generosity of sponsors.

Knowledge in your pocket

The second advance is the availability of knowledge “in hand”. Back in the day, we had pocket reference cards that summarised the essentials. But these were limited – how much could you stash in your white lab coat pocket?  

 Image: DALL-e

Today, more knowledge than you could possibly keep in your head, or any pocketbook is easily stored in your smartphone, in an eBook or an app, or available through live connection to online knowledge sources. [Lab coats are increasingly a relic, their sleeves an infection transmission hazard and professional styles of dress, at least in SA, are now less formal].

Automated interpretation

The next advance was automated interpretation from bedside ECG machines. In the 1970s, feature-based rules were encoded in them to generate diagnostic reports, which appear on the printout as a clinical aid. Debate exists about the accuracy and reliability of these systems, but until recently they were the pinnacle of computerised decision support. Our UCT teachers and many others however regard them as unreliable.

A 12-lead ECG with automated reporting diagnosing normal findings. Image: North Dakota State University

Along comes AI

Machine learning-based approaches are repowering the ECG.

Large labelled data sets have been used to train neural networks to do tasks that humans do, like rhythm classification, but at scale, quicker and better.

Second, these AI algorithms can identify subtle patterns, potentially in parts of the ECG not used in traditional analysis, enabling new capabilities:

  • Exceeding human capabilities by identifying conditions (e.g., heart muscle dysfunction) that are not reliably recognised by expert human readers, and
  • Identifying individuals at risk of developing disease not yet detected with standard tests.

AI ECGs can detect heart muscle dysfunction and disease, rhythm disturbances (such as atrial fibrillation (AF)), heart valve narrowing, high blood pressure in the lungs (pulmonary hypertension), even non-cardiac conditions such as liver disease (cirrhosis) and blood electrolyte disturbance (high potassium levels).

How the artificial intelligence (AI) ECG performs these new predictive tasks is not fully understood. It probably reflects disease processes at a cellular level, affecting the activity of sodium and potassium ions moving in and out of cells, that cause subtle ECG changes that precede findings from tests like echocardiograms (ultrasound pictures of the heart) which show physical function.

Beagles and Eagles

Cardiologists love acronyms – every big cardiology study must have one. The EAGLE study recruited over 22,000 patients and showed that when routine ECGs were ordered by primary care practitioners, the availability of an AI ECG result increased the detection of left ventricular (heart muscle) dysfunction by one-third.

Image: DALL-e

In the BEAGLE study, 669,000 patients without known atrial fibrillation (“AF” – the most common chronic heart rhythm disturbance) were screened with an AI ECG algorithm then invited to use a wearable monitor for a month. Patients whose AI ECG showed “silent” AF had a 5-fold increased likelihood of clinical AF detection in the next month. 

The AI ECG has also been shown to reflect drug treatment effect. AI ECG scores detect hypertrophic cardiomyopathy, an important form of heart muscle disease, and fall in response to treatment.

Patients do their own ECGs

The Apple Watch and other consumer devices allows patients to do their own, limited yet accurate ECGs which can detect atrial fibrillation. Taking this a stage further, the Mayo Clinic Center for Digital Health created an app to collect Apple Watch ECG recordings and with machine learning methods trained the watch ECG system to reliably identify left ventricular (heart muscle) dysfunction (area under the curve 0.89). An unaided cardiologist can’t do this.

Video-based teaching is more fun

The CASSA course is excellent but it doesn’t have Ray the Moose (RRRAIMUS) – a character who appears in entertaining, video-based approaches used to teach ECGs and other clinical content. Probably more effective as a teaching method than sitting through a lecture with powerpoint slides.

Ray the moose, a memory aid. Image: SKETCHY

RRRAIMUS stands for Rate (atrial), Rate (ventricular), Rhythm, Axis, Intervals, Morphology, Unique electrolyte changes, and Summary

Things have changed – but some things shouldn’t and haven’t

Some things don’t, and shouldn’t change – for example, the requirement for the systematic approach that the UCT/CASSA faculty advocate. Quick takes on an ECG can be misleading, and dangerous. A good diagnosis needs the full story, in context.

The ECG does not tell the full story. A clever AI may do the task but does not do the job. Clinical context and patient history, laboratory tests and imaging hold the information needed for accurate diagnosis. All relevant findings need to be available, in real time, integrated into clinical workflows and acted upon in a timely fashion. An algorithm that is predictive is not the same as a clinical intervention.

One more change

To get it right more often, in good time, and make a difference (i.e. make the best possible clinical decision), the ability to easily share the ECG with knowledgeable, experienced colleagues might be the biggest advance of all.

Many older studies showed that peer to peer decision support is both more common and more highly valued than computer-based decision support, an often-overlooked way that clinicians (through “hallway consultations” – and now remotely) meet their information needs. Fax machines once helped do this job, multimodal technologies take it to the next level. WhatsApp, though not officially sanctioned as a medium for exchange of private health information, is a powerful platform for clinical interactions in routine everyday use in SA, offering huge value to clinical teams at almost zero cost.


It’s clear, things have changed. For decades the ECG  has been critical for diagnosing cardiac and some forms of non-cardiac disease but with AI, integration into clinical workflow, use of communication and collaboration platforms, and new teaching methods, the ECG’s role as a power tool in the pockets of clinicians continues to grow, 100 years after it first arrived. The ECG can detect disease, predict disease, and monitor treatment effect. It’s inexpensive, available in non-medical environments and even to consumers!

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top