Neuron selection for decoding dexterous finger movements.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-machine interfaces (BMI) use neural activity to reconstruct kinematics.
• Non-task-related neurons can cause overfitting and reduce decoding accuracy in BMI models.
• Existing neuron selection methods do not adequately address overfitting issues.

Purpose of the Study:

• To introduce a novel method, cross-model validation, for selecting task-relevant neurons in BMI.
• To improve the generalization and accuracy of neural decoding models.
• To enhance the ability of BMI to distinguish between different movements.

Main Methods:

• Developed a cross-model validation technique to rank neuron importance.
• Evaluated neuron importance based on model generalization to correct and incorrect movement data.
• Compared decoding accuracy using cross-model validation selected neurons versus random selection.

Main Results:

• Cross-model validation effectively identifies neurons crucial for distinguishing movements.
• Selecting neurons with cross-model validation significantly improves decoding accuracy.
• Decoding accuracy increased up to 2.5 times (44%) compared to random neuron selection.

Conclusions:

• Neuron selection is critical for accurate BMI decoding.
• Cross-model validation is a powerful tool for assessing individual neuron utility in decoding.
• This method enhances BMI performance by focusing on informative neural signals.