Extended Data Fig. 1: Technological and computational design underlying the BSI.

a, Photographs reporting the geometry and features of the WIMAGINE implant, including 64 platinum-iridium (90:10) electrodes with 4 mm x 4.5 mm pitch (in antero-posterior and medio-lateral axes respectively). Two external antennas are embedded within the implant. The first antenna powers the implanted electronics through inductive coupling at high frequency (HF, 13.56 MHz) while the second ultrahigh frequency antenna (UHF, 402-405 MHz) transfers the recorded signals outside the body. b, Two external antennas embedded in a personalized 3D-printed headset power the implant and recover the streamed signals that are then transferred to a base station. This base station manages the powering of the implants, synchronization and conditioning of the raw data. c, A decoding pipeline computes temporal, spectral and spatial features embedded in the ECoG signals related to the intention to move. These features are then uploaded into the decoding algorithm that decodes the attempts to move the lower limbs based on a tailored, recursive exponentially weighted Markov-switching multi-linear model algorithm²⁶. This algorithm is a mixture of multilinear experts’ algorithm integrating a Hidden Markov Model (HMM) classifier, called gating, and a set of independent regression models, called experts. The gating classifier predicts the joint that is intended to be mobilized (i.e. hip, knee or ankle on each side) as well as resting state, while each expert is dedicated to predicting the direction and relative amplitude of the intended movement. When updating is allowed, every 15 s, the coefficients of both linear regressions (β^gate, b^gate, β^expert, b^expert) are updated through recursive partial least square along with the coefficients of the transition matrix T corresponding to the number of transitions between each states during this 15s period (i.e. 150 new transitions). To support the production of standing and walking, the outputs of the model are encoded into updates of joint-specific stimulation programs that are constrained within pre-established functional ranges of amplitudes. d, A tailored, medical-grade software sends these updates to the implanted pulse generator through a chain of wireless communication systems, eventually delivering the stimulation through a paddle array implanted epidurally over the lumbosacral spinal cord.