Prediction of peak oxygen uptake from resting 12-lead electrocardiograms using deep learned representations in competitive athletes
G Bohus, E Csulak, K Szabo, L Batai, L Szabo, H Vago, D Becker, B Merkely, N SydoAbstract
Background/Introduction
Cardiopulmonary exercise testing (CPET) is the gold standard for measuring peak oxygen uptake (VO2max), but it requires a special technical and personnel background. Deep-learning models based on 12-lead resting electrocardiograms (ECG) have demonstrated potential for predicting VO2max in clinical populations, but their effectiveness in competitive athletes has not yet been investigated.
Purpose
To evaluate whether pretrained convolutional ECG representations (PCLR) can predict VO2max from a 10-second 12-lead resting ECG in competitive athletes, and to compare the predictive value of ECG-derived features with clinical covariates alone.
Methods
Athletes underwent detailed sports cardiology screening (questionnaire, resting ECG, blood tests, echocardiography, CPET, and body composition analysis). The VO2max was determined by CPET. Resting ECGs were processed through a validated deep learning network to extract 320 features representing cardiac electrical activity. Two predictive approaches were compared: (1) an ECG-only model using 50 deep learning features selected by statistical methods on training data only; and (2) a Basic model including 46 clinical features: training history, laboratory, echocardiographic, and body composition parameters. Data were split into subject-disjoint training (n=578) and test (n=144) sets with 10-fold subject-aware cross-validation. Performance was assessed using mean absolute error (MAE) and Pearson correlation (r) with bootstrap 95% confidence intervals.
Results
We studied 722 competitive athletes (age: 18.3±4.7 years; 75% male; training hours/week: 16.4±6.0; VO2max 52.6±6.5 mL/kg/min) with both 12-lead resting ECG and CPET on record. On the independent test set, the Basic clinical model predicted VO2max with MAE 4.08 mL/kg/min (95% CI 3.54-4.63) and r=0.657 (95% CI 0.541-0.752). The ECG-only model achieved MAE 4.26 (95% CI 3.65-4.88) and r=0.612 (95% CI 0.525-0.689) (Figure 1.). Cross-validation confirmed both findings. More than 93% of predictions fell within the Bland-Altman limits of agreement. The strongest clinical predictors were sex, lean body mass, height, age, and lung function.
Conclusion(s)
According to our analyses, VO2max can be predicted using only deep-learned features of the 12-lead resting ECG in competitive athletes. Our ECG-only model achieved predictive accuracy similar to the Basic model, highlighting the role of resting ECG patterns in the physiologic sports adaptation in athletes. These findings support the feasibility of ECG-based VO₂max estimation in athletic populations as a non-exercise alternative for assessing cardiorespiratory fitness.figure 1