Artificial intelligence capable of detecting left ventricular hypertrophy: pushing the limits of the electrocardiogram?

This editorial refers to ‘Comparing the performance of artificial intelligence and conventional diagnosis criteria for detecting left ventricular hypertrophy using electrocardiography’, by J.-M. Kwon et al., on pages 412--419. As cardiologists interpreting electrocardiograms (ECGs), we routinely record the presence of left ventricular hypertrophy (LVH) using a variety of manual rule-based measurements and calculations. It makes sense that the ECG should be a good measure of LVH—electrical voltage and conduction time should correlate well with left ventricular mass. Assessing voltage and the duration of impulse conduction is an essential step in comprehensive ECG interpretation and seems worthwhile. After all, LVH is an independent predictor of cardiovascular morbidity and mortality in hypertensive patients1 and early recognition and treatment can improve clinical outcomes.2 Three of the most common manually applied diagnostic methods for LVH detection are the Sokolow–Lyon criteria,3 Romhilt–Estes point-score system,4 and the sex-specific Cornell voltage criterion.5 Despite the routine application of these criteria, their diagnostic performance is surprisingly lacklustre. Thus, their ability to detect or screen for early onset LVH is limited. Even attempts to improve diagnostic accuracy by incorporating various factors (e.g. body mass index, age, and sex) have been inadequate.6,7 These criteria were derived from small samples of patients with unclear demographics most of whom had underlying cardiac disease—along with a ‘gold standard’ for LVH that varied among the studies and would be considered inadequate today (M-mode echocardiography, heart weight measured in small number of samples at autopsy, or even using surrogates such as the presence of aortic stenosis, coarctation of the aorta, or hypertension). Similarly, the statistical approaches for defining these criteria were state of the art more than 50 years ago but seem antiquated today. There is a clear need to bring more contemporary approaches to solving this important diagnostic dilemma. Advances in the field of deep learning have unlocked new promising capabilities in medicine, especially ECG. Our group and others have demonstrated that artificial intelligence-enabled ECG (AI-ECG) algorithms have been shown to predict a patient’s chronological age and sex8 as well as underlying cardiac function.9 In a world with ever-advancing data availability and new smartphone-based electrodes, the possibilities for an AI-ECG algorithm to display diagnostic aptitude are only beginning to emerge. In this issue of Europace, Kwon et al.10 developed and validated an AI-ECG algorithm to detect LVH. This retrospective cohort of over 20 000 patients from two hospitals who had an ECG and echocardiogram performed around the same time. They took an interesting methodologic approach, comparing several deep and machine learning processes. They found that algorithm that combined raw ECG data and structured data (i.e. demographics, biometrics, and structured data from the ECG), the so called ensemble neural network (ENN), performed better than either the convolution neural network (CNN) performed on raw ECG signal or a deep neural network (DNN) that examined only structured data. This ENN performed well, demonstrating an area under the receiver operating characteristic curve in the external validation sample of 0.868.7 However, it is important to note that the best model required a combination of a DNN that included structured demographic/biometric information and the CNN that examined the raw ECG data. Although the CNN on the raw signal significantly outperformed conventional LVH criteria manually applied by cardiologists (i.e. Romhilt–Estes point-score system, Cornell voltage criteria, and Sokolow–Lyon criteria), there remained room for improvement in its performance, even with relatively straightforward additional information. Currently, convolutional models applied to large, high quality datasets is considered the state of the art for using raw signal data in binary classification problems, but they too have limits. As data scientists are often fond of saying, ‘a black box is not a magic box’. The ability of an AI-ECG algorithm to detect LVH better than currently devised manual criteria and computer-generated algorithms is another step forward in the field of deep learning in medicine. The growing body of literature in this field underscores the incredible power of a simple, inexpensive, largely operator-independent, and non-invasive tool that remains yet to be fully known. An AI-ECG algorithm that improves the performance of existing methods only enhances the care patients receive by improving diagnostic screening, identifying optimal treatment, and predicting prognoses. In the current medical landscape with ubiquitous digital data stored in electronic health records, perhaps we can move beyond the idea that diagnostic tests should be interpreted in isolation—the current paper seems to support this concept. Drawing from the medical record, automated algorithms could assess pre-test probability of a given disease state and could fine tune an algorithm’s sensitivity and specificity for detection of a disease state of interest for an individual patient. Similarly, the results of a test could be constrained by known patient-specific comorbidities or demographic features. We are not far away from a time in which an AI-ECG algorithm, perhaps while being bolstered by data from other sources, will fully interpret a standard 12-lead ECG with much better accuracy than is currently accepted, and ultimately improve clinical workflow, triage, and delivery of healthcare. Conflict of interest: none declared. The opinions expressed in this article are not necessarily those of the Editors of Europace or of the European Society of Cardiology.