Real-time Instantaneous Phase Estimation Using a Deep Dual-Branch Complex Neural Network.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Accurate estimation of neural oscillation phase is vital for brain-interfacing technologies like brain-computer interfaces (BCIs) and neuromodulation.
• Traditional phase estimation methods, such as the Hilbert transform using Discrete Fourier Transform (DFT), introduce phase lags due to their reliance on past and present data.

Purpose of the Study:

• To develop a deep learning algorithm for accurate instantaneous phase estimation of neural oscillations.
• To design an algorithm suitable for deployment on resource-constrained edge devices like FPGAs for real-time applications.

Main Methods:

• A novel deep learning algorithm employing a dual-branch structure, inspired by the complex wavelet transform, was developed.
• Discrete Cosine Transform (DCT) layers were utilized to extract latent representations for signal components, generating a pseudo-complex signal.
• The algorithm was designed for efficiency, targeting deployment on portable edge devices with limited computational power.

Main Results:

• The proposed deep learning model demonstrated a significant improvement in phase estimation accuracy.
• Achieved a 40.3% increase in accuracy compared to the endpoint-corrected Hilbert Transform (ecHT) method.
• Showed a 9.2% improvement over conventional deep learning architectures.

Conclusions:

• The developed deep learning algorithm offers a more accurate and efficient method for instantaneous phase estimation in neural signals.
• This proof-of-principle work validates the potential for real-time phase estimation in neuromodulation and other BCI applications.
• The algorithm's suitability for edge devices opens possibilities for advanced, portable brain-interfacing systems.