A Task-Driven Adversarial Channel Selection Method for Binary Classification Based on Magnetocardiography.

Area of Science:

• Biomedical Engineering
• Cardiology
• Signal Processing

Background:

• Increasing sensor density in magnetocardiography (MCG) arrays leads to data redundancy and resource inefficiency.
• Existing channel selection methods often fail to balance computational efficiency, accuracy, and scalability, necessitating redesigns with changing conditions or sensor malfunctions.

Purpose of the Study:

• To introduce a task-driven adversarial channel selection method for binary classification of MCG signals.
• To enhance the efficiency, practicality, and scalability of MCG systems by optimizing channel selection.

Main Methods:

• Developed a task-driven adversarial channel selection approach for MCG signal binary classification.
• Employed a group-wise search with a heuristic algorithm to determine optimal channel combinations.
• Defined an objective function to maximize the difference between classification accuracy and cosine similarity of selected channels.

Main Results:

• Successfully reduced the number of channels from 36 to 5 without compromising classification performance on an MCG dataset.
• Outperformed the hybrid sequential forward search method in accuracy and channel reduction.
• Demonstrated superior scalability compared to hybrid sequential forward search and Pearson-rank methods.

Conclusions:

• The proposed method effectively balances computational cost and classification accuracy.
• Achieved improved scalability of selected channel combinations, enhancing MCG system practicality.
• Offers a robust solution for efficient and reliable MCG data analysis.