Background and Aims: Athletes like footballers can suffer from abnormal heart rhythms called cardiac arrhythmias. The electrocardiogram (ECG) reveals these arrhythmias as rate- or morphology-based arrhythmias (depolarisation or repolarisation). Disentangled representation learning of ECGs offers an unsupervised approach to arrhythmia classification in athletes, but requires inductive bias about the data. The inductive bias influences the VAE loss-weight profile, which differs across arrhythmia types. Suboptimal loss-weight profiles risk misclassifying life-threatening arrhythmias in footballers. We leverage homoscedastic uncertainty estimation to optimise pathology-specific (rate-, depolarisation-, and repolarisation-based abnormalities) VAE loss weightings for arrhythmia classification in footballers.
Methods: We use a β-total correlation (TC) VAE, whose loss function decomposes into the reconstruction, TC, classification, and mutual information contrastive loss. We categorise the 30 most important arrhythmias from the PhysioNet Challenge 21 dataset (n=88,253) into a baseline dataset and three pathology-specific datasets (rate, depolarisation, and repolarisation) for the β-TC-VAE training. The β-TC-VAE loss weights are optimised using homoscedastic uncertainty estimation for each dataset. We evaluate the β-TC-VAE on the PF12RED dataset (n=163) of professional footballers, comparing classification performance using the area under the receiver operating characteristic curve, the area under the precision-recall curve, and the F1-score.
Results: Initial results indicate superior performance of the β-TC-VAE loss-weight optimised model compared to the VAE trained with the baseline dataset when applied to arrhythmias in footballers. Depolarisation- and repolarisation-based arrhythmias yielded higher reconstruction loss weights than rate-based arrhythmias, reflecting the morphological complexity over rate abnormalities.
Conclusion: This study demonstrates how homoscedastic uncertainty estimation optimises pathology-specific VAE loss weights for arrhythmia classification in footballers. The higher reconstruction loss weights for depolarisation- and repolarisation-based arrhythmias reflect their greater morphological complexity compared to rate abnormalities. Future work should include explainability methods to facilitate clinical implementation and prospective validation to enhance the distinction between physiological and pathological conditions in athletes.