Generalizable Finger Movement Decoding from Intracranial Recordings Across Static and Dynamic Actions

Reliable decoding of finger movements requires brain-computer interfaces (BCIs) that generalize across the diverse static and dynamic actions performed in daily life. Static postures and dynamic movement sequences rely on partially distinct neural representations, and the balance between these components varies widely across tasks. Yet most BCI decoding pipelines are optimized for a narrow task domain, limiting their ability to transfer to new scenarios. Here we show that standard intracranial decoding approaches show notable performance drops when applied to previously unseen finger tasks with different static-dynamic compositions, and that generalization critically depends on multiple elements of the decoding pipeline. High gamma features and short temporal windows yield the most robust cross task transfer. Although nonlinear models perform best when examples of all tasks are represented during training, linear decoders generalize better to novel, particularly dynamic, actions. We identify a consistent static-dynamic structure across tasks that may provide a useful basis for improving generalization. Finally, we show that anatomical heterogeneity between tasks also constrains generalization. Together, these results establish design principles for BCIs that generalize reliably across the rich repertoire of human finger movements.