Key result
Cycle-patching and domain-balance tracking improves phonocardiogram classification scores by ~5% over standard methods.
Why the study?
Training deep learning models for phonocardiogram classification is challenged by data heterogeneity and noisy inherited labels from multiple heart cycles in early valvular disease stages.
Answers are grounded in this paper and cite their sources.
A novel deep learning approach using cycle-patching and domain-balanced training improves phonocardiogram classification performance in heterogeneous datasets with noisy labels.
Phonocardiograms (PCGs) can be intelligently processed to diagnose heart valve-related diseases earlier than their symptoms manifest using advanced deep learning models trained using large datasets of heart sound measurements. As large sets of measurements (data) cannot be obtained using a single type of sensor or acquired from a particular measurement site on the chest of a subject or in a strictly controlled environment, the resulting datasets are nonhomogeneous. Even though training is done on a heart cycle basis, a record consisting of several continuous heart cycles is given a single class label, normal or abnormal, and each cycle then inherits this label. Inherited class labels may be incorrect when measurements are obtained from subjects having early stages of valvular diseases. Training with heterogeneous measurements, along with possible wrong labels, has been ignored in the literature. This article proposes constructing tensors from multiple cycles to mitigate the effects of noisy labels (NLs) and using domain-balanced training with a tracking method to ensure that all measurements contribute to model training and address heterogeneity. Vertical and horizontal patching are proposed to extract meaningful features from the constructed input tensors before feeding them into a focal modulation transformer and then fusing features for final classification through a majority voting process. Comprehensive experiments demonstrate that the proposed method achieves average score improvements of 19%, 3%, and 5% over centralized training, domain-balanced training, and the closest state-of-the-art method, respectively. These results demonstrate the effectiveness of the proposed approach for improving binary classification performance under data heterogeneity and noisy-label conditions. The source code is available at https://github.com/EbrahimAli1989/cycle-patching-pcg-classification.git
No takes yet. Share an insight, caveat, or question.
Nehary et al. (2026) studied Heart valve-related diseases. Cycle-patching and domain-balance with tracking vs. Centralized training, domain-balanced training, and state-of-the-art methods was evaluated on Binary classification performance. A proposed cycle-patching and domain-balance tracking method for phonocardiogram classification improved average scores by 5% compared to the closest state-of-the-art method.
Synapse has enriched 10 related papers. Consider them for comparative context: