رومينا ادهمي

Networks for decoding movement parameters from signals recorded in area 2 of the primary somatosensory cortex (S1), a region whose neurons respond to proprioceptive feedback from joints an‎d muscles during both active an‎d passive movements an‎d play an important role in the representation of proprioception. Since the neural activity of this area can serve as a basis for providing artificial proprioceptive an‎d tactile feedback in brain–computer interface (BCI) systems, eva‎luating the decoding capacity of movement-related information from this region is considered a fundamental step toward the development of advanced neuroprosthetic systems. For this purpose, neural data obtained from two adult rhesus monkeys were collected over four intracortical recording sessions while performing a Center-Out task under both active an‎d passive movement conditions. These data were analyzed in order to predict two categories of movement variables, including kinematic parameters (position, velocity, an‎d joint angle) an‎d kinetic parameters (force an‎d moment), in four target directions (0°, 90°, 180°, an‎d 270°). Three deep neural network architectures were designed an‎d implemented, including a ClassificationNet for movement direction classification, a RegressionNet for regression of movement parameters, an‎d a MultiTaskNet for simultaneous classification an‎d regression. After performing signal preprocessing an‎d filtering, the models were trained an‎d eva‎luated using metrics such as accuracy, F1-score, area under the curve (AUC), root mean square error (RMSE), an‎d coefficient of determination (R²). The results indicated that decoding accuracy in active movements was significantly higher than in passive movements, reflecting a greater richness of information under active conditions. Moreover, the RegressionNet model achieved the best performance in predicting joint angle, while the MultiTaskNet model provided balanced, stable, an‎d reliable results, particularly at 0° an‎d 180°. The findings of this study demonstrate that area 2 of the primary somatosensory cortex has a high capacity for decoding movement-related information an‎d can serve as a reliable source for providing proprioceptive feedback in BCI systems. The proposed models can provide a foundation for future studies aimed at improving generalizability, utilizing multimodal data, an‎d developing intelligent rehabilitation systems based on brain–computer interfaces.

يادگيري عميق , رابط بين مغز و كامپيوتر

Deep learning , Brain computer interface

Romina Adhami

Mohammadreza Daliri