The rough but fascinating road to estimate peak exercise oxygen uptake by resting electrocardiogram-based deep learning

This editorial refers to ‘Deep learned representations of the resting 12-lead electrocardiogram to predict $\dot{\mathbf{V}}{\mathbf{O}}_{\mathbf{2}}$ at peak exercise’, by S. Khurshid et al., https://doi.org/10.1093/eurjpc/zwad321.

Maximal exercise performance depends on the activity of multiorgan systems, genetic substrate, and environmental factors whose interaction comes together into a single final measurement that is oxygen uptake (VO₂) at peak exercise, i.e. the product of O₂ delivery times peripheral extraction throughout diffusion and mitochondrial O₂ utilization.¹ Its direct measure by gas exchange analysis with cardiopulmonary exercise testing (CPET) has gained value across decades because of the indisputable ability to predict the risk of cardiovascular events in the apparently healthy population^2,3 and, even more, in patients with cardiovascular disease (CV) diseases.⁴

This is because a single number stands for all the biological complexity to orchestrate the physiological chain of O₂ transport and uptake in matching the external (air) to internal (mitochondria) ventilation. Therefore, VO₂ peak measured by breath-by-breath analysis, averaging the latest 30 s of maximal exercise, is the gold standard measure of exercise capacity.¹

Insights on pathophysiology and high reproducibility represent main strengths compared with alternative tests, such as the 6 min walk test.^5,6 Directly measured VO₂ also provides accuracy compared with functional estimates of maximal performance such as workload and metabolic equivalents.

Despite the many advantages of measuring VO₂ peak, it requires dedicated software, careful equipment calibration, and instructed personnel, all factors that may impact in the daily clinical practice.¹

In the present issue of the Jounal, Khurshid et al.⁷ provide a large population-based analysis of subjects performing CPET, looking at the 12-lead electrocardiogram (ECG) prediction ability to estimate VO₂ peak. Prediction models were validated internally (Massachusetts General Hospital, MGH) and externally at Brigham and Woman Hospital (BWH). Then, the association between ECG estimated peak VO₂ and incident disease (atrial fibrillation, myocardial infarction, heart failure, and all-cause death) was independently tested in 84 718 primary care subjects with an ECG recorded within 3 years prior to the event-driven follow-up. Three learning models were developed: the Basic (regression model ‘controlling’ VO₂ for age, sex, body mass index, and type of exercise); the Basic + ECG intervals (heart rate, PR, and QT intervals and QRS duration); and the Deep ECG-VO₂. Estimated VO₂ peak by each of these methods correlated with true CPET-derived measure, but the Deep ECG-VO₂ yielded the highest accuracy both at MGH and BWH sites. This model also performed better compared with the Wasserman, Jones, and FRIEND VO₂ predictive reference equations. Estimation of a low exercise performance by the Deep ECG VO₂ model strongly predicted cardiovascular and all-cause mortality when tested in the independent ambulatory primary care setting.

Findings are highly contributory focusing, for the first time, on the role of resting ECG, over the previous attempts to derive VO₂ peak estimates by cardiac-related variables at rest.^8–10 The artificial intelligence methodology incorporating the intervals’ assessment of standard ECG trough deep learning models is sound providing intriguing implications for the modern cardiology times, mainly when addressing population-based studies. The rationale is likewise strong and quite surprisingly under-appreciated in the past, considering that electrical activity is basically involved in the cardiac dynamics ensuring the progressive enhancement in O₂ delivery under exercise metabolic requests. The sample of investigated populations is large and proportionally balanced in those at risk but apparently healthy and the subjects with CV diseases. Ultimately, this approach can be the ideal way to estimate the severity of functional limitation in frail patients, cancer survivors, and those with a physical impediment to perform a maximal exercise test.

Nonetheless, correlation analyses between Deep VO₂ peak estimation and CPET-measured indicate an average range of error quite wide, i.e. 5.5 mL/min/kg, corresponding to a more than one class difference in the Weber classification, which means a significant negative or positive change in cardiovascular event prediction. Along with these important clues, the time lag between resting ECG recording and CPET evaluation, though short, is questionable. Intuitively, the assessment of ECG tracings before the CPET sessions should have avoided any unanticipated change between the two tests.

To prove model robustness, the ECG VO₂ peak model was tested also in the single and wider group without main ECG abnormalities, such as left ventricular hypertrophy, left and right bundle branch block, atrial fibrillation, and left ventricular ejection fraction <50%. This, however, does not guarantee full accuracy in patients with an abnormal ECG at presentation, traditionally referred to exercise laboratories for CPET evaluation because of cardiac and ventilatory limitation. Conceivably, the implications of estimate of VO₂ peak in this population are more impactful than in the general population, and future studies should clarify this issue.

Although female representation was up to 40%, the average age was quite low, and information lacks on the elderly population who most likely may benefit more from peak VO₂ deep learning estimation. Also, a dedicated analysis for individuals who could not undergo a maximal test due to their physical inability or frailty condition would have been highly indicated.

As for any newly proposed surrogate measure, the large number of collected data and the documented power of the Deep ECG-VO₂ to predict events are certainly a good start for next applications in the clinical decision-making process. However, the demonstration that a new approach has value against the existing standard requires the validation and demonstration of reproducibility across laboratories.

Overall, the work by Khurshid et al.⁷ is meritorious because it challenges our dogmas on VO₂ peak measurements showing that there is a fascinating road for its estimation that, although rough, can be ‘deeply’ uncovered by resting ECG learning.

Author notes

The opinions expressed in this article are not necessarily those of the Editors of the European Journal of Preventive Cardiology or of the European Society of Cardiology.

Conflict of interest: None declared.