Decoding with limited neural data: a mixture of time-warped trajectory models for directional reaches

Neuroprosthetic devices can help paralyzed patients perform daily tasks by decoding their intentions from neural signals like electromyograms (EMGs). To improve accuracy, decoders use trajectory models that consider how the body naturally moves. This study enhances trajectory models by incorporating likely movement targets (identified by gaze) and variations in reach speeds to improve decoding accuracy, especially when only a few EMG signals are available. The algorithm uses a mixture of extended Kalman filters (EKFs) to combine insights related to movement speed variation and probabilistic target knowledge. It was tested using EMGs and eye movements to decode hand position during 3D reaching tasks.

• 1
  Incorporating gaze-based target information significantly improves decoding accuracy compared to using EMG data alone, especially at simulated injury levels.
• 2
  A time-warped trajectory model (mTWTM) that accounts for variations in movement speed outperforms models that do not, particularly for faster reaches.
• 3
  The mixture of Kalman filters with targets (mKFT) demonstrates better generalization to novel targets compared to the mixture of trajectory models (MTM) when a fully specified training set is not available.