Improved Grip Force Prediction Using a Loss Function that Penalizes Reward Related Neural Information

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Neural activity in sensorimotor cortices correlates with movement and non-sensorimotor variables like reward.
• Brain-Machine Interfaces (BMIs) aim to decode neural signals for device control.
• Optimizing prediction accuracy in BMIs is crucial for effective application.

Purpose of the Study:

• To compare the performance of an artificial neural network (ANN) for offline grip force prediction using two distinct loss functions.
• To evaluate the impact of a modified reward-penalized mean squared error (RP_MSE) loss function against the standard mean squared error (MSE).
• To identify brain regions where the RP_MSE loss function yields improved prediction accuracy.

Main Methods:

• An artificial neural network (ANN) was trained for offline grip force prediction.
• Two loss functions were employed: standard mean squared error (MSE) and a novel reward-penalized mean squared error (RP_MSE).
• Prediction performance was assessed across dorsal premotor cortex (PMd), primary motor cortex (M1), and primary somatosensory cortex (S1).

Main Results:

• The ANN demonstrated significantly improved grip force prediction performance when using the RP_MSE loss function.
• Performance enhancement was observed in three key sensorimotor brain regions: PMd, M1, and S1.
• The improvement in prediction accuracy under RP_MSE was approximately 6% compared to MSE.

Conclusions:

• The RP_MSE loss function is a more effective approach for optimizing ANN-based grip force prediction in BMIs.
• Penalizing the correlation between reward and grip force improves neural decoding accuracy in sensorimotor cortices.
• This finding has implications for developing more robust and accurate Brain Machine Interfaces.