Stable and artifact-resistant decoding of 3D hand trajectories from ECoG signals using the generalized additive model.

Abstract

The key criterion for reliability of brain-computer interface (BCI) devices is their stability and robustness in natural environments in the presence of spurious signals and artifacts. To improve stability and robustness, a generalized additive model (GAM) is proposed for BCI decoder identification. Together with partial least squares (PLS), GAM can be applied to treat high-dimensional data and it is compatible with real-time applications. For evaluation of prediction quality, along with standard criteria such as Pearson correlation, root mean square error (RMSE), mean absolute error (MAE), additional criteria, mean absolute differential error (MADE) and dynamic time warping (DTW) distance, are chosen. These criteria reflect the smoothness and dissimilarity of the predicted and observed signals in the presence of phase desynchronization. The efficiency of the GAM-PLS model is tested on the publicly available database of simultaneous recordings of the continuous three-dimensional hand trajectories and epidural electrocorticogram signals of the Japanese macaque. GAM-PLS outperforms the generic PLS and improves the evaluation criteria: 22% (Pearson correlation), 8% (RMSE), 13% (MAE), 31% (MADE), 20% (DTW). Motor-related BCIs are systems to improve the quality of life of individuals with severe motor disabilities. The improvement of the reliability of the BCI decoder is an important step toward real-life applications of BCI technologies.