Background and Aims: Cardiac arrhythmias are abnormal heart rhythms that can occur in athletes. These arrhythmias can be classified into their cardiac cycle (systolic, diastolic, or both), electrophysiological phase (depolarisation or repolarisation), or anatomical origin (supraventricular, conduction disturbance, or ventricular). Shaping neural network (NN) architectures to specific types of arrhythmias is an emerging artificial intelligence (AI) research direction. We leverage Neural Architecture Search (NAS), an automated machine learning technique, to explore NN architecture optimisation for arrhythmia classification. We aim to explore NAS patterns and temporal and spatial ECG feature preferences for more targeted AI-based arrhythmia classification approaches.
Methods: We propose NAS-driven arrhythmia classification by cardiac cycle, electrophysiological phase, and anatomical origin. The NAS search space consisted of AutoFormer, a chain-structured search space which combines convolutional operations and self-attention blocks. We then applied the Differentiable ArchiTecture Search (DARTS) strategy to the search space. We used the \num{30} most important arrhythmias from the PhysioNet Challenge 21 dataset ($n = \num{88253}$) and categorised the arrhythmias into their cycle, phase, and origin. We evaluated the NNs on the PF12RED arrhythmia dataset ($n = \num{163}$) of professional football players, comparing performances with and without the NAS optimisation using the F1-score and the area under the receiver operating characteristic curve.
Results: Initial results indicate superior performance of the NAS-optimised model compared to the baseline network when applied to sports-related arrhythmias. The NAS preferred converging towards smaller architectures for all three arrhythmia classifications, suggesting efficient feature extraction from athletic electrocardiograms.
Conclusion: This study explores how NAS can optimise NNs for sports-related cardiac arrhythmia classifications. The convergence towards smaller architectures suggests that the NNs can effectively encode the most relevant diagnostic information from electrocardiogram signals. Future work should include prospective validation on interpretability to facilitate clinical implementation and enhance the distinction between physiological and pathological conditions in athletes.