Biophysically constrained learning from healthy neural priors improves motor decoding in pathological cortical networks.

High-performance brain-computer interfaces (BCIs) rely on stable neural representations, yet brain functional impairments inherently disrupt canonical encoding, creating a fundamental decoding paradox. Conventional data-driven approaches struggle to bridge this gap because of pathological data scarcity and heterogeneity. Here, we propose to leverage healthy neural priors and introduce Pathology-Aware Consensus Training (PACT), a paradigm embedding biophysical constraints into healthy neural recordings to guide neural decoding. In a gesture classification task using high-density electrocorticography from patients with gliomas infiltrating the motor cortex, PACT improved the accuracy by more than 10%, effectively restoring the distorted semantic geometry of motor commands and realigning them with healthy canonical representations (correlation increase: r = 0.66 to r = 0.94). Mechanistically, PACT induced an emergent "inverse weighting" strategy: The model autonomously suppresses high-power pathological noise while reanchoring spatial attention to functional compensatory regions. This study provides a proof of concept for improving decoding in compromised networks by harnessing healthy neural dynamics, laying the theoretical foundation for future BCI deployments in clinical populations.