A CNN model embedded with local feature knowledge and its application to time-varying signal classification

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Signal Processing

Background:

• Accurate classification of multi-channel time-varying signals, particularly with imbalanced datasets, remains a challenge.
• Existing methods often struggle to effectively model both global and local signal features.
• Cardiovascular disease diagnosis using electrocardiogram (ECG) signals requires robust feature extraction and classification techniques.

Purpose of the Study:

• To propose a novel convolutional neural network (CNN) architecture for local prior feature embedding and imbalanced dataset modeling in multi-channel time-varying signal classification.
• To enhance the classification accuracy and generalization capabilities of signal analysis models.
• To improve the modeling of imbalanced datasets by leveraging local feature information.

Main Methods:

• A novel CNN architecture comprising parallel single-channel feature extraction, multi-channel feature integration, and local feature embedding units.
• Utilized dynamic clustering and sliding window for typical local feature set generation.
• Employed external embedding with sliding window and dynamic time warping (DTW) for local feature similarity measurement.

Main Results:

• The proposed method effectively extracts and represents both global and local signal features.
• Demonstrated improved modeling of imbalanced datasets by strengthening the role of prior local features.
• Achieved significantly improved accuracy and generalization in cardiovascular disease classification using 12-lead ECG signals.

Conclusions:

• The developed CNN methodology offers a robust approach for multi-channel time-varying signal classification, especially with imbalanced data.
• The integration of local prior features and advanced similarity measurement enhances diagnostic performance.
• The technique shows significant promise for applications in medical signal analysis and disease diagnosis.