Machine learning-based high-frequency neuronal spike reconstruction from low-frequency and low-sampling-rate...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Multichannel electrophysiological recordings are crucial for understanding brain function.
• Handling large datasets and hardware limitations pose challenges in neuronal recording.
• Thermal tissue damage is an unavoidable issue with current hardware.

Purpose of the Study:

• To develop a machine learning-based method for reconstructing high-frequency neuronal spikes from subsampled low-frequency signals.
• To overcome hardware limitations and data handling challenges in multichannel electrophysiological recordings.
• To enable more comprehensive analysis and control of brain functions.

Main Methods:

• Applied a transformer machine learning model to neuronal data.
• Utilized subsampled low-frequency band signals to reconstruct high-frequency neuronal spikes.
• Trained and validated the model on both in vitro and in vivo mouse brain data.

Main Results:

• The machine learning model accurately estimated high-frequency neuronal spike information, including spike timing and waveform, from x8 downsampled data.
• Network connectivity could be reasonably inferred from the reconstructed signals.
• The method demonstrated effectiveness on both in vitro and in vivo datasets.

Conclusions:

• Machine learning-based reconstruction offers a viable solution for reducing data size in electrophysiological recordings.
• This approach is compatible with existing multichannel recording hardware.
• The technique facilitates broader bandwidth neuronal signal acquisition, enabling advanced brain function analysis and control.