Denoising and decoding spontaneous vagus nerve recordings with machine learning.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Peripheral nerve stimulation via neural interfaces aids in managing conditions like epilepsy and depression.
• Effective neuromodulation requires efficient extraction of neural data, often challenged by low signal-to-noise ratios (SNR) and non-stationary noise.
• Machine learning (ML) shows promise but requires adaptation for biomedical signal processing tasks.

Purpose of the Study:

• To adapt and evaluate ML algorithms for unsupervised denoising of neural recordings.
• To compare ML-based denoising with traditional bandpass filtering for vagus nerve activity.
• To assess the efficacy of ML in preserving relevant neural features for closed-loop neuromodulation.

Main Methods:

• Applied bandpass filtering and two novel ML algorithms to in-vivo vagus nerve recordings.
• Utilized a variational autoencoder (VAE) for unsupervised denoising.
• Compared denoising performance by extracting respiratory afferent activity.

Main Results:

• The variational autoencoder (VAE) model demonstrated superior performance compared to bandpass filtering.
• VAE-based denoising showed better correlation with respiratory activity.
• Performance was validated using cross-correlation, Mean Squared Error (MSE), and accuracy in determining respiratory rate.

Conclusions:

• ML algorithms, particularly VAEs, can effectively denoise neural recordings while preserving essential biological signals.
• These ML approaches offer a promising avenue for enhancing the efficacy of implantable neuromodulation devices.
• Further development of ML for neural signal processing is crucial for advancing closed-loop biomedical applications.