Enhanced heart sound anomaly detection via WCOS: a semi-supervised framework integrating wavelet, autoencoder and SVM

Area of Science:

• Data Mining
• Machine Learning
• Biomedical Signal Processing

Background:

• Anomaly detection is crucial in data mining, often involving imbalanced datasets.
• Semi-supervised methods are efficient but struggle with noisy, uneven data, impacting accuracy.
• Detecting conditions like congenital heart disease requires precise anomaly identification.

Purpose of the Study:

• To enhance the accuracy of semi-supervised anomaly detection on imbalanced datasets.
• To propose a novel method, WCOS, that addresses noise and data imbalance.
• To improve the detection of subtle anomalies, such as abnormal heart sounds.

Main Methods:

• Developed a new semi-supervised anomaly detection method (WCOS).
• Integrated wavelet reconstruction, convolutional autoencoder (CAE), and one-class classification support vector machine (OCSVM).
• Utilized semi-supervised clustering as a core component of the WCOS framework.

Main Results:

• WCOS demonstrated superior anomaly detection accuracy compared to existing methods.
• The method effectively filters noise, enhancing the distinction of abnormal samples.
• Evaluated on real datasets, WCOS showed significant improvements in AUC standard deviation under noise.

Conclusions:

• WCOS offers a robust solution for anomaly detection in imbalanced and noisy datasets.
• The combined approach of wavelet reconstruction, CAE, and OCSVM enhances detection precision.
• The proposed method significantly outperforms traditional approaches like OCSVM, WR-OCSVM, and CAE-OCSVM.