Decoding Overlapping Lower-Limb Afferent Pathways from Human Epidural Spinal Recordings

Restoration of dynamic motor function following neurological injury increasingly relies on adaptive neuroprostheses, which require real-time sensory feedback to continuously adjust to a user's physical state. However, it remains unknown whether distinct afferent activity can even be decoded from highly overlapping, volume-conducted epidural fields. This challenge is particularly pronounced in the lumbosacral enlargement, where common fibular (CFN) and tibial nerve (TN) afferents converge extensively, producing highly similar cord dorsum potential (CDP) topographies. Here, we demonstrate for the first time in humans that clinical-grade lumbosacral epidural paddle arrays capture sufficient fine-scale spatiotemporal structure to decode these overlapping inputs. Using a 32-contact array and peripheral nerve stimulation, we constructed a 42-dimensional feature space capturing distributed amplitudes, field geometry, and waveform morphology. A support vector machine decoded four distinct afferent classes (left and right CFN and TN) with a median accuracy of 90.9% ± 0.3%. Shapley Additive Explanations revealed decoding was driven by contact-level voltage patterns and temporal waveform complexity, while the geometric features contributed minimally. These afferent-specific signatures persisted even at sub-motor threshold stimulation intensities. By utilizing standard clinical arrays, this approach provides a pathway toward rapid deployment of interpretable closed-loop neuromodulation, avoiding the surgical risks of penetrating or peripheral interfaces.