Unsupervised decoder initialization for brain-machine interfaces using neural state space dynamics

One of the key elements in the design of neuro-motor Brain-Machine Interfaces (BMIs) is the neural decoder design. In a biomimetic approach, the decoder is typically trained from concurrent recordings of neural and kinematic or motor imagery data. The non-availability of the latter data imposes a practical problem for patients with lost motor functions. An alternative approach is a biofeedback approach in which subjects are encouraged to `learn' an arbitrary mapping between neural activity and the external end effector. In this work, we propose an unsupervised decoder initialization scheme to be used in the biofeedback approach that alleviates the need for synchronized kinematic or motor imagery data for decoder training. The approach is totally unsupervised in that the recorded neural activity is directly used as training data for a decoder designed to provide `desirable' features in the decoded control signal. The decoder is trained from `spontaneous' neural data when the BMI subject is not engaged in any behavioral task, and we demonstrate its ability to generalize to neural data collected when the subject is in a different behavioral state.