Deep Neural Oracles for Short-Window Optimized Compressed Sensing of Biosignals

Area of Science:

• Signal Processing
• Machine Learning
• Biomedical Engineering

Background:

• Sparse signal recovery is crucial in data compression and analysis.
• Traditional methods involve separate support and coefficient estimation phases.
• Existing techniques face challenges in efficiency and self-assessment of recovery quality.

Purpose of the Study:

• To develop an integrated encoding-decoding scheme for sparse signal recovery.
• To enhance recovery capabilities for biological signals like ECG and EEG.
• To introduce a self-assessment mechanism for reconstructed signal quality.

Main Methods:

• A deep neural network-based oracle estimates signal support.
• The oracle is trained jointly with a linear mapping encoder.
• Coefficient estimation is performed via pseudo-inversion using the oracle's output.
• The scheme is applied to electrocardiogram (ECG) and electroencephalogram (EEG) signals.

Main Results:

• Achieved state-of-the-art signal recovery capabilities.
• Enabled extremely low-complexity encoders for biological signals.
• Demonstrated accurate self-assessment of reconstructed signal quality.
• Achieved a compression factor of 2.67 for ECG/EEG signals with low computational cost.

Conclusions:

• The proposed oracle-based recovery scheme offers efficient and accurate sparse signal reconstruction.
• The self-assessment feature provides unique insights into reconstruction reliability.
• This approach paves the way for improved, application-specific signal processing.