Improving movement trajectory models through time-warping and multi-sensor fusion
To restore movement to a person with a paralyzed arm, the intended movement of the user must be estimated from biological signals under voluntary control using a brain machine interface (BMI). State of the art approaches use Bayesian statistics to combine the probabilistic predictions of a trajectory model – a model of how the state of the device (e.g. position and velocity of the arm) evolves over time with the predictions of an observation model – a mapping from the state to the biological control signals. The trajectory model may incorporate any prior assumptions we have about the nature of the desired movement.
There has been recent interest in designing more accurate trajectory models; taking advantage of the directional nature of reaching movements in particular has led to improved decoding performance. However, incorporating too many parameters into the model may lead to deficits in terms of generalization. For example, if a system is designed to produce extremely stereotyped movements to a specific set of targets it may perform poorly when a novel target is introduced.
Another aspect of reaching which has not previously been addressed by standard trajectory models is the desired movement speed. If a person wants to move more slowly or quickly than normal the dynamics of the reach will be warped – stretched or compressed in time. These effects cannot be accounted for using a linear trajectory model.
The goal of this work was to design a trajectory model that would improve decoding for BMIs with an application to reaching. We incorporated two features that prominently influence the dynamics of natural reach: the movement speed and the target location. Our approach incorporates uncertain target information which is obtained from eye-tracking. We estimate the average speed throughout the course of the reach, and implement a nonlinear model with the flexibility to follow the dynamics of the reach more accurately. The model generalizes well to new regions of the workspace for which there is no training data, and across a broad range of reaching dynamics to widely spaced targets in three dimensions.