Non-negative constrained dictionary learning for compressed sensing of ECG signals

Area of Science:

• Signal Processing
• Biomedical Engineering
• Machine Learning

Background:

• Compressed sensing (CS) relies on overcomplete dictionaries for signal reconstruction.
• Traditional dictionary learning methods (ℓ0, ℓ1 norms) can yield negatively correlated atoms, hindering reconstruction.
• Suboptimal dictionary atom correlation leads to inefficient signal representation in CS.

Purpose of the Study:

• To propose a Non-Negative Constrained Dictionary Learning (NCDL) algorithm.
• To enhance the reconstruction performance of CS for electrocardiogram (ECG) signals.
• To address limitations of existing dictionary learning methods in CS.

Main Methods:

• NCDL algorithm with non-negative constraints on encoding coefficients.
• Alternating direction method of multipliers for sparse solutions.
• Block coordinate descent for dictionary updates.
• Integration of a penalty term to refine sparse coding.

Main Results:

• NCDL demonstrated superior ECG signal reconstruction compared to existing CS methods.
• Achieved a compression ratio (CR) of approximately 2.78 at a 9% PRD1.
• Outperformed methods like optimal direction, k-means SVD, and online dictionary learning within a CR range of 1.05 to 2.78.

Conclusions:

• The NCDL algorithm effectively improves CS reconstruction performance for ECG signals.
• Non-negative constraints on dictionary atoms enhance signal representation and sparsity.
• NCDL offers a promising new direction for CS applications in biomedical signal processing.