Comparative Evaluation of Neural Manifolds for Force Decoding from Intracortical Neural Activity

Brain-machine interfaces (BMIs) connect the brain and external devices, aiding those with motor disabilities like spinal cord injuries (SCI). Decoding neural activity is challenging due to its high dimensionality, noise, and variability. Neural manifold techniques help by representing neural activity in a stable, low-dimensional space. This study compares two neural manifold techniques, principal component analysis (PCA) and demixed principal component analysis (dPCA) for decoding force parameters from intracortical neural activity in BMIs. Results indicate that PCA achieves a higher decoding performance, with a correlation coefficient (R) of 0.87 and a coefficient of determination (${\mathrm {R}}^{2}$) of 0.75, compared to dPCA’s R of 0.86 and $\mathrm{R}^{2}$ of 0.74. Despite this marginal difference, dPCA excels in separating neural components related to specific movement parameters, offering a more detailed representation of neural dynamics. This work highlights dPCA’s potential to enhance BMI accuracy and reliability by leveraging its taskspecific capabilities, which could significantly improve movement decoding in complex BMI applications.