A Swallowing Decoder Based on Deep Transfer Learning: AlexNet Classification of the Intracranial Electrocorticogram.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-machine interfaces (BMIs) are crucial for assistive technologies, particularly for swallowing disorders.
• Electrocorticogram (ECoG) signal decoding is essential for developing effective swallowing BMIs.
• Current methods often involve complex signal processing, such as extracting high-frequency power bands.

Purpose of the Study:

• To investigate the efficacy of deep transfer learning using raw electrocorticogram (ECoG) signals for swallowing detection.
• To compare the performance of decoding raw ECoG signals against processed ECoG power bands.
• To assess the feasibility of using pre-trained deep learning models (AlexNet) with visually transformed neural data.

Main Methods:

• ECoG signals were recorded from eight epilepsy patients during swallowing tasks.
• Raw ECoG signals and specific frequency bands were converted into time-series images.
• Deep transfer learning with AlexNet was applied to classify swallowing events using these images.

Main Results:

• The model achieved 74.01% accuracy using high-gamma band (75-150 Hz) ECoG power.
• Decoding accuracy using raw ECoG signals reached 76.95%, comparable to processed data.
• Sensitivity and specificity were 82.51% and 95.38% respectively, for the processed data.

Conclusions:

• Deep transfer learning is effective for decoding swallowing movements from ECoG signals.
• Raw ECoG signals can be used directly for transfer learning, eliminating the need for conventional high-gamma band extraction.
• This approach simplifies BMI development for swallowing assistance.