cnn-snn-eeg-imagined-speech

name	cnn-snn-eeg-imagined-speech
description	Hybrid CNN-SNN architecture for EEG-based imagined speech decoding. First integration of spiking neural networks into imagined speech BCI, achieving 80.13% accuracy on BCI Competition III benchmark. Activation: imagined speech, EEG decoding, CNN-SNN hybrid, spike-based BCI, neuromorphic BCI
tags	["neuroscience","BCI","SNN","EEG","imagined speech","hybrid architecture"]

Overview

First study to integrate Spiking Neural Networks (SNNs) into EEG-based imagined speech decoding, addressing the challenge of non-stationary, low-amplitude EEG signals through biologically-inspired spike-based temporal dynamics.

Core Methodology

Hybrid CNN-SNN Pipeline

1. CNN Stage: Extracts temporal representations from raw EEG signals
2. SNN Stage: Performs biologically-inspired temporal classification using spike-based mechanisms
3. Integration: Leverages event-driven firing mechanisms of biological neurons

Key Innovations

• Exploits spike-based temporal dynamics not captured by traditional deep learning
• Event-driven firing mechanisms model biological neuron behavior
• First application of SNNs to imagined speech decoding

Performance

• Dataset: 2020 BCI Competition III benchmark
• Accuracy: 80.13% (surpassing previous best of 70.19%)
• Evaluation: Comparable evaluation settings with existing methods

Applications

• Brain-computer interfaces for communication restoration
• Neuromorphic BCI applications
• Next-generation assistive technologies for speech impairments

Implementation Notes

• CNN handles feature extraction from non-stationary EEG signals
• SNN captures temporal dynamics through spike-based processing
• Hybrid approach outperforms pure CNN or pure SNN architectures

Pitfalls

• EEG signals are non-stationary and highly variable across subjects
• Imagined speech produces weaker signals than actual speech
• Traditional ML/DL methods fail to exploit spike-based temporal dynamics

References

• arXiv:2607.03844 (July 2026)
• Authors: Fatima Shalhoub, Mariam Al Mawla, Kabalan Chaccour, Iván López-Espejo, Hoda Fares
• Accepted to IEEE EMBC 2026