In-depth education and ECG Applied Cardiology

Laurent Fiorina, Mina AIT says, Jack Cartier Hospital, Massey

Until recently, ECGs could be mistakenly considered as part of the arsenal of past cardiologists. When it comes to diagnostics, while we’ve seen impressive advances in imaging, the ECG, discovered by Willem Anthoven in 1895, hasn’t changed much in almost a century. However, it remains a cornerstone of diagnostics in medicine and especially cardiology, with more than 300 million ECGs performed annually worldwide.(1).

The current ECG revolution is a combination of two major advances: the introduction of the latest artificial intelligence technology, including deep learning (DL), and the advent of wearable or portable attachments. . One of the most advanced artificial intelligence technologies, deep learning or “deep learning” is one of the most advanced artificial intelligence technologies currently in use. It was popularized by Google in 2015 thanks to the victory of AlphaGo, an algorithm that defeated the champion of Go of Go, a game known as the most complex. Similarly, the use of facial recognition by DL algorithms has become widespread. One of the first proofs of the effectiveness of this technology for medical use was made for the diagnosis of diabetic retinopathy (2). In early 2014, engineers in collaboration with cardiologists applied it to the ECG. This involves training a DL algorithm in a large ECG database with accurate diagnostics so that the algorithm is able to detect different ECG waves and their distortions. Thus, the first device of its kind to obtain certification was Holter ECG analysis with CE marking from Cardiologs in 2016, then FDA in 2017, showing an analysis time saving without compromising diagnostic performance (3). Then came the Apple Watch for diagnostics, including the Cardia device in 2017 and the ECG watch in 2018, and other accessories, including the EKO ECG stethoscope to detect low LVEF. These algorithms were originally created for ECGs with one or more leads (holters, portable devices, and watches), which allow many diagnoses, especially for rhythmic, but also for 12-lead ECGs, where they prove their superiority over conventional algorithms. Has (4). An essential prerequisite for the use and training of these algorithms is the digitization of ECGs and databases. In fact, this technology requires sufficiently well-organized, well-annotated data and accurately linking to the elements we want to “learn” in the algorithm. Also, techniques exist to interrogate the algorithm so that a posterior to know the elements (figures) indicating its conclusions. Currently, in the first large-scale study published in Two Types of Application Diagnosis Nature, AY Hannun et al. (5) compare a DL algorithm with the diagnosis of major arrhythmias of 91,232 ECGs (1 lead) by 6 cardiologists and show superiority over the latter. The effectiveness of the algorithm here was certainly impressive, but the cardiologists had a lower “sensitivity” to detect arrhythmias (FA 71%, BAV 73%, VT 65%). Regarding the 12-lead ECG diagnosis, AH Ribeiro et al. (6) reported interesting results in 827 tracing with 89.5%, 100% and 100% positive predictive values ​​for BBD, BBG and FA, respectively. S. Al-Jaiti and his team (7) tested their algorithm on the ability to diagnose an acute coronary syndrome on an ECG and increased the sensitivity by 37% without losing the negative predictive value. The range of classic criteria for left ventricular hypertrophy is well known. The challenge of diagnosing structural heart disease, such as hypertrophic cardiomyopathy (HCM), has been addressed by WY Ko et al. (8) were met by proven HCM in a population of 612 patients with a sensitivity and specificity of 87 and 90% and 12 788 controlling patients. ECG is considered normal by experts when maintaining similar performance. Furthermore, the algorithms tested in a derivation (D1) revealed a sensitivity and specificity of 83 and 81%. The idea here is that a DL algorithm trained on a large number of examples will have more efficient “eyes” to distinguish subtle changes that cannot be detected even by an expert human. ECG as a biomarker The term biomarker is applied to the ECG in relation to the possibility of finding a characteristic signature of a disease, a diagnosis or a response to treatment. This weak ECG signal, ignored or invisible to the naked eye, may be associated with cardiac or extracardiac disorders, present or future. ZI Attia (9), P. Bachtiger (10) and their teams have succeeded in training an LD algorithm to predict lower LVEF (<35%) from the training base of 44,959 Mayo Clinic patients where each 12 D ECG corresponds to an ultrasound. This algorithm was later validated by other studies and used by the EKO device (recently marketed) that combines an enlarged stethoscope with a bipole to allow a 1D ECG to be recorded. Perhaps less intuitively, algorithms have been developed to detect biological imbalances such as hyperglycemia. (11) with still incomplete results (sensitivity and specificity at 87 and 85%). Similarly, 1,638,546 ECGs from the Mayo Clinic matched serum potassium levels during recording, enabling CD Galloway et al. (12) to create an LD algorithm for predicting hyperkalemia with a sensitivity and specificity of 90% and 63%. So DL can predict one condition at a time, but also provide future pathology with adequate data available with longitudinal follow-up and the pathology lends itself to it. Through a neural network training on 1.6 million ECGs of 430,000 patients followed between 1984 and 2019. Raghunath et al (13) have shown this. The principle is to predict from 12D sinus ECG if patient will develop AF in 1 year, which seems possible with sensitivity and specificity of 69 and 81%. We can imagine the application of such algorithms in screening or post-stroke assessment. Deep learning at home Finally, DL algorithms, with the help of devices like ECG clocks, make it possible to increase the number of remote diagnoses outside a medical environment. It offers the possibility of remote observation, Covid 19 has become important by epidemic and medical population, with more frequent and more precise observation. The diagnosis of AF with ECG clock is now well known. Another use case is the monitoring of QT intervals, evaluated with good results in the QT-logs study (14).

Please note that this site is reserved for health professionals for regulatory reasons.

To see more, register for free.

If you are already registered,
Enter:

If you are not yet registered on the site,
Free registration:

Leave a Comment