Neural Correlates of error processing during grasping with invasive brain-machine interfaces

Brain-machine interfaces (BMIs) may generate more errors than those encountered during normal motor control. Thus, they provide an opportunity to investigate neural correlates of error processing. Characterizing neural correlates of error processing may, in turn, provide a tool for on-line correction of the errors that are made by the interface. We investigated neural correlates of error processing during BMI experiments in which monkeys controlled an animated hand on the screen to touch a ball by moving their own fingers. Short movement segments that were consistently toward or away from the target were labeled accordingly and used to train a classifier to differentiate between correct and erroneous movements based on the neural activity. The results indicate that despite the limited number of labeled segments and active neurons in the studied data, the classifier achieved a classification rate of 68% on testing. The full receiver operating curve (ROC) has been estimated and indicates that even when the false alarm is restricted to 5%, the classifier can detect 36% of the erroneous movements. Better results are expected when using more data, especially as more challenging grasping tasks are performed. Such a classifier could be used to improve the performance of BMIs by detecting and correcting erroneous movements.