DeepFingerNet Predicts Finger Trajectory From ECoG Measurements

Decoding fine motor movements, such as finger trajectories, from brain signals poses a critical challenge for brain-computer interface (BCI) systems, particularly when employing electrocorticography (ECoG) because of its inherent signal complexity and inter-subject variability. Although traditional machine learning and deep learning methods have advanced coarse motor control, they often fall short in capturing and generalizing. To address these limitations, we propose DeepFingerNet– a novel architecture based on nested U-Nets designed specifically for ECoG-based finger trajectory prediction. DeepFingerNet enhances feature extraction and fusion by leveraging nested skip connections, which effectively integrate low-level spatial details with high-level abstract representations. We employ a hybrid loss function, combining cosine similarity and mean squared error, to optimize the direction and magnitude of the decoded trajectories. Evaluations on two publicly available ECoG datasets demonstrate the model’s effectiveness, achieving correlation coefficients of 0.69 (a 17% improvement over baseline methods) on BCI Competition IV Dataset 4 and 0.54 (a 5% improvement) on the Stanford University dataset, respectively. These results underscore DeepFingerNet’s potential to significantly improve the reliability and precision of fine finger predictions from ECoG measurements, thereby supporting advanced applications in neuroprosthetics and human-machine interaction.