Deep-learning based prediction of peak oxygen uptake in athletes using 2D echocardiographic videos

Abstract

Abstract Funding Acknowledgements Type of funding sources: Public grant(s) – EU funding. Main funding source(s): Project no. RRF-2.3.1-21-2022-00004 (MILAB) has been implemented with the support provided by the European Union. Introduction Cardiopulmonary exercise testing (CPET)-derived peak oxygen uptake (VO2/kg) is a well-established parameter of exercise capacity allowing the quantification of athletic performance. Although VO2/kg is mainly influenced by anthropometric and demographic factors, several studies demonstrated strong associations between resting echocardiography-based measures and VO2/kg. Artificial intelligence could incorporate information from both features, thus enabling a more accurate prediction of exercise capacity in athletes. Aim Accordingly, we aimed to implement a deep-learning (DL) model that uses 2D echocardiography (2DE)-based apical 4-chamber view videos on top of the anthropometric features (age, sex, body surface area [BSA]) to predict VO2/kg and then assess the model’s performance in a large cohort of athletes. Methods We retrospectively identified 422 competitive athletes (15.4±7.3 training hours/week) who underwent resting 2DE evaluation and then CPET to determine VO2/kg (52.7 ± 7.7 mL/kg/min). To predict VO2/kg values, we trained a deep neural network that can process both modalities of the inputs (i.e. 2DE videos and anthropometric data such as age, sex and BSA) simultaneously (Figure 1). We applied 5-fold cross-validation and used mean squared error (MSE), mean absolute error (MAE), and R squared (R2) metrics to measure our model’s performance. Then, we compared the results with linear regression that was trained using only the 3 anthropometric factors (age, gender, BSA). Additionally, after finalization of the DL-based model, we prospectively recruited further 25 competitive athletes with both 2DE and CPET performed to validate our model. Results Using 2DE videos, our DL-based model was able to achieve an accurate prediction of VO2/kg with an MSE of 35.27, MAE of 4.62, and an R2 of 0.393. In comparison, the linear regression model using only anthropometric factors had worse predictive performance in all metrics with an MSE of 40.51, MAE of 4.88, and R2 of 0.303. In addition, we compared the predictive performance of the DL-based and the linear regression models by their respective squared error values using the Wilcoxon test. Our DL-based model had a significantly better performance compared to the linear regression model (Wilcoxon p = 0.006). In the prospective dataset, our DL-based model achieved an MSE of 16.69, MAE of 3.42, and an R2 of 0.169, whereas the linear regression model was inferior with an MSE of 25.43, MAE of 4.51, and an R2 of −0.268. The DL-based model showed a significantly better performance (Wilcoxon p<0.001). Conclusions Using our DL-based approach on our large athlete database, we were able to implement and prospectively validate a model that incorporated 2DE videos to predict VO2/kg more accurately compared to using anthropometric factors alone. DL techniques may advance sports medicine by personalized monitorization of training phases and accurate prediction of athletic performance.