مقالات پذیرفته شده در نهمین کنگره بین المللی زیست پزشکی
A Supervised Contrastive Learning Approach for ECG Label Noise Learning
A Supervised Contrastive Learning Approach for ECG Label Noise Learning
Yixue Li,1,*Yuchen Wang,2Yongjian Li,3Yiheng Sun,4Chen Meng,5
1. School of Control Science and Engineering, Shandong University 2. School of Control Science and Engineering, Shandong University 3. School of Control Science and Engineering, Shandong University 4. School of Control Science and Engineering, Shandong University
Introduction: With the widespread use of wearable devices, it has become increasingly easy to obtain electrocardiogram (ECG) signals. Manual annotation is time-consuming, labor-intensive, and prone to errors. The learning effectiveness of deep learning models depends on high-quality databases, and the presence of label noise can affect the learning effectiveness of the model. This study aims to reduce the impact of label noise on the model's learning of ECG classification.
Methods: A label noise learning (LNL) method based on supervised contrastive learning is proposed, which consists of two parts: feature extraction and label detection. For feature extraction, a projection layer is added after the general model to extract features, and supervised contrastive learning is used to make features from the same annotated samples more similar. For label detection, the true category and label reliability of a sample are accurately determined by quantifying the similarity between the features of a single sample and those of other samples. During training, the system automatically retains cleanly labeled samples and uses model prediction results to correct identified noise labels.
Results: At a noise rate of 40%, the proposed method achieved accuracy improvements of 29.87% and 32.11% on the arrhythmia database and atrial fibrillation database, respectively, compared to methods that do not use LNL. Compared to state-of-the-art methods, the improvements were 5.68% and 1%, respectively. Additionally, the method demonstrated high accuracy in identifying clean labels during training, with an accuracy rate exceeding 90%, thereby validating its effectiveness.
Conclusion: This study proposes a supervised contrastive learning-based label noise method that effectively improves the accuracy and robustness of ECG classification.
Conclusion: This study proposes a supervised contrastive learning-based label noise method that effectively improves the accuracy and robustness of ECG classification.