Leveraging advanced machine learning techniques to develop robust movement decoders

Spinal cord injury (SCI) is one of the major causes of paralysis worldwide. Since the communication between the brain and the spinal circuits is interrupted, people affected by SCI present a lack of movement and sensation below the level of the injury. For these patients, the restoration of arm and hand movements is their priority. One of the strategies is Epidural Electrical Stimulation. It consists in the application of current in the epidural space of the spinal cord. It has been proved that EES recruits sensory afferent fibres within the dorsal roots. Motor patterns are then generated through synaptic excitation of motor neurons. The simplest EES protocol is the continuous (or tonic) one. It consists in the tonic application of current pulses with predetermined constant parameters to a single region of the spinal cord. Another possible stimulation protocol is the phasic, in which the parameters remain constant and predefined (as in the tonic situation), but the current is applied through bursts elicited in particular phases of the movements (i.e., rich and pull in an object manipulation task or leg extension and leg flexion in a locomotion task). Finally, continuous decoding stimulation refers to a strategy in which stimulation site and parameters change accordingly to the intended muscle activity, decoded from cortical signal. The work presented here is aimed at designing a new neural decoder to predict the intended muscle activity in order to implement it in a continuous stimulation decoding protocol. Furthermore, the effect of the EES on neural dynamics (of premotor and motor cortex) was analysed in order to understand why stimulation caused the previously used decoders to fail. Experiments were executed on a female macaque monkey (Macaca Fascicularis), in which two 64-channels microelectrode Utah arrays (Blackrock Microsystems, US), eight intramuscular bipolar electrodes for EMG acquisition and a customized spinal implant (epidural electrode, described in the next section) were implanted. The intramuscular electrodes were implanted into eight muscles of the arm, forearm and hand: deltoid flexor carpi radialis (FCR), extensor carpi radialis (ECR), flexor digitorum superficialis (FDS) and abductor pollicis brevis (ABP). The spinal implant is made by seven independent electrodes embedded in a matrix of soft silicon. The active sites are distributed along the vertical axis in order to selectively recruit specific subgroup of muscles. The effect of stimulation through the electrode’s active sites was characterized by recruitment experiments performed on the anesthetized animal. Through a stimulus generator (Multi Channel Systems, Germany), current-controlled, charge-balanced biphasic pulses are delivered. During the experiments, the animal had to reach and then pull the end effector of a robotic arm (KUKA, Germany) in nine different positions. While performing the task, four kinds of data were acquired: (1) the intramuscular electromyographic activity (Tucker-Davis Technologies acquisition system, US), (2) the neural activity from premotor and motor cortex (Blackrock Microsystems acquisition system, US), (3) arm kinematics (VICON motion capture system, UK) and (4) pull and grip forces on the end effector. The trials were done in different stimulation conditions: (1) baseline, where the stimulation was absent, (2) tonic, in which the stimulation was constant and continuous in a certain active site and with predefined stimulation parameters, and (3) phasic, in which the stimulation consisted in current bursts during reach phases or pull phases of the arm movement. Three different type of decoding strategies were chosen to be tested: Wiener Cascade (a combination of a linear filter and a cascade non-linearity), Kalman Filter and Long Short-Term Memory Network (LSTM – a special type of recurrent neural network). Performances were evaluated using the fraction of variance accounted for. The neural response at phasic stimulation was found characterized by a peak at 398 ms (mean among neurons, SD = 117 ms). After this maximum, the firing rate decreases but does not return to the initial baseline. PCA was used to analyse the tonic condition. After gaussian smoothing of the firing rate, the neural trajectories (plot of the first three principal components) were extracted. It is possible to notice a shift of the neural trajectories from the baseline condition to the stimulation one, by maintaining more or less the same dynamics. Independent Component Analysis on firing rate was carried out for both tonic stimulation condition and phasic stimulation condition. Some of the ICs, at visual inspection, presented differences in the stimulation condition with respect to the baseline condition. Regarding the decoding performances, LSTM performed much better than the other two decoders, with a mean R^2 of 0.75. Training the network on baseline trials and testing it on stimulation trials resulted in poor decoding accuracy for the trials in which the stimulation highly affected the muscle activity. However, analysing a recording day in which the stimulation had little or no effect, the decoding accuracy of the LSTM was comparable to that of baseline decoding. The same was not true for Wiener Cascade and Kalman Filter, suggesting that the LSTM can explicitly ignore the effect of stimulation on brain activity, treating it simply as noise