Continuous decoding direct neural interface which uses a markov mixture of experts

A method of continuous decoding of motion for a direct neural interface which acquires neural signals using a plurality of electrodes, wherein a command variable, y(t), is estimated which represents kinematic motion variables, using a mixture of experts, each expert being associated with one of the hidden states of a HMM model which models an observation variable represented by an observation vector, x(t), where the HMM model comprises at least one active state and one idle state, said method comprising:

• a calibration phase for estimating parameters of the HMM model and of a linear prediction parameter of each expert;
• pre-treatment of neural signals, said treatment including a time-frequency transformation over a moving observation window, in order to obtain a sequence of observation vectors x[1:t];
• estimating the mixture coefficients (gk(t),k=1, . . . , K) of the various experts from the probability of said sequence of observation vectors (P(x[1:t])) and from the probability of the respective states associated with these experts, given said sequence (P(zk(t)=1|x[1:t]));
• estimating by each expert, Ek, said motion variable from the running observation vector (x(t)) and the linear prediction parameter of said expert;
• combining estimates of the various experts using mixing coefficients in order to provide an estimate of the command variable;
• and
• wherein the mixture coefficient gk(t) associated with the expert Ek is obtained by
• g
• k
• ⁡
• (
• t
• )
• =
• P
• ⁡
• (
• z
• k
• ,
• x
• ⁡
• [
• 1
• ⁢
• :
• ⁢
• ⁢
• t
• ]
• )
• P
• ⁡
• (
• x
• ⁡
• [
• 1
• ⁢
• :
• ⁢
• ⁢
• t
• ]
• )
• where (P(x[1:t]) is the observation probability of the sequence of observation vectors and (P(zk,x[1:t]) is the state probability associated with the expert, given the sequence observed, with the probabilities (P(x[1:t]) and (P(zk,x[1:t]) being obtained using a forward algorithm.