Real-time low latency estimation of brain rhythms with deep neural networks

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Neurofeedback and brain-computer interfaces (BCIs) require real-time interpretation of brain activity for effective closed-loop communication.
• Minimizing the delay between neural events and feedback is crucial for enhancing interaction efficacy.
• Existing methods for isolating narrow-band brain signals introduce fundamental delays, necessitating more efficient approaches.

Purpose of the Study:

• To explore the utility of modern time series forecasting neural networks for low-latency extraction of brain rhythm parameters.
• To develop and evaluate novel, efficient approaches for tracking brain rhythm phase and envelope for instantaneous brain interaction.
• To investigate the potential of neural networks in compensating for delays in brain signal processing.

Main Methods:

• Tested five distinct neural network architectures for forecasting synthetic electroencephalography (EEG) rhythms.
• Identified the strongest forecasting architecture and trained it to simultaneously filter and forecast EEG data.
• Compared the performance of the selected neural network against state-of-the-art techniques (cFIR, Kalman filter, Conv-TasNet) using synthetic and real EEG data from 25 subjects.

Main Results:

• The Temporal Convolutional Network (TCN) demonstrated superior forecasting performance, achieving >90% rhythm envelope correlation with <10 ms effective delay and <20° circular standard deviation of phase estimates.
• The TCN exhibited stability against noise perturbations.
• When trained for filtering and prediction, the TCN outperformed cFIR and Kalman filter methods, and matched the performance of the larger Conv-TasNet architecture.

Conclusions:

• Demonstrated the efficacy of a neural network approach, specifically TCNs, for low-latency narrow-band filtering of brain activity signals.
• The proposed TCN-based framework enhances the effectiveness of brain-state dependent paradigms in various applications.
• This forecasting framework provides a valuable tool for investigating EEG signal predictability and understanding brain information processing.