| CPC G06F 3/015 (2013.01) [A61B 5/30 (2021.01); A61B 5/316 (2021.01); G06F 18/2113 (2023.01); G06F 18/214 (2023.01); G06F 18/295 (2023.01); G06N 7/01 (2023.01); G06N 20/00 (2019.01); G06N 20/20 (2019.01)] | 12 Claims |
|
1. A method of calibrating a direct neural interface that will receive a plurality of electrophysiological signals acquired using a plurality of sensors, during a plurality of observation windows associated with observation times, and provide command signals for one or more effectors configured to produce a trajectory or trajectories replicating a corresponding trajectory or trajectories of an imagined movement corresponding to one or more of the plurality of electrophysiological signals, said plurality of electrophysiological signals being preprocessed to obtain an observation tensor (Xt) at each observation time (t), changes in the observation tensor being modelled by a hidden Markov Model (HMM), said direct neural interface using a mixture of a plurality K of experts, each expert (Ek) being associated with a hidden state (k) of the HMM model and being defined by a multi-linear expert predictive model, comprising:
receiving the electrophysiological signals;
processing the electrophysiological signals to produce the observation tensor (Xt); and
calibrating during a plurality of calibration phases, at predetermined instants of said trajectory, each calibration phase (u) corresponding to a plurality (ΔL) of successive observation times, and making use of a tensor Xu representing an observation tensor at said plurality (ΔL) of successive times, a tensor Yu representing a set of command tensors at these same times and a matrix Zu, giving states of the HMM model at these same times, wherein each calibration phase comprises:
a step (a) in which tensors Xu Yu, are extracted using the matrix Zu, observation tensors Xuk and command tensors Yuk, k=1, . . . , K, relative to the different states (k) of the HMM model;
a step (b) in which the observation tensor is input to the plurality K of experts and, for, each expert Ek, tensors Xuk and Yuk are calculated corresponding to tensors Xuk and Yuk respectively, after being centred and normalised, then a covariance tensor of tensor Xuk and a cross-covariance tensors of tensors Xuk and Yuk being modified by adding covariance and cross-covariance tensors Xu-1k and Yu-1k respectively derived from a previous calibration phase, weighted by a forget factor, λk;
a step (c) of using a multivariate regression of partial least squares with exponential recursive weighting (REW-NPLS) regression, starting from covariance and cross-covariance tensors modified in the previous step to generate prediction coefficient tensors and a prediction bias tensor which are input to the linear predictive expert models for training the linear predictive expert models, thereby updating the linear predictive expert models;
a step (d) in which the tensor Xu corresponding to the tensor Xu, after being centred and normalised, the covariance tensor of tensor Xu and the cross-covariance tensor of
 u and Zu are calculated, the covariance tensor of tensor Xu and cross-covariance tensor of  u and Zu being modified by adding to them the covariance tensor of  u-1 and the cross-covariance tensor of  u-1 and Zu-1, derived from the previous calibration phase, weighted by a forget factor λ;a step (e) of training a second multi-linear predictive model using an REW-NPLS regression starting from covariance and cross-covariance tensors modified in the previous step to generate a prediction coefficients tensor and a bias vector which are input to the second multi-linear predictive model, thereby updating the second multi-linear predictive model to give a state vector (zt) of the HMM model as a function of the centred and normalised input tensor,
 u, components of the state vector providing possibilities that the direct neural interface is in each of the different states k=1, . . . ,K at an observation time, and generating mixture coefficients (νkt, k=1, . . . ,K) of different experts using the second multi-linear predictive model; anda step (f) of providing an estimate (Ŷt) of a control tensor representing the command signals.
|