Unidirectional LSTM-based 3D reaching trajectory prediction using selected magnetometer-only SQUID-MEG channels

Visual induced motor kinematic activities play a significant role in many aspects, such as decoding and controlling activities in rehabilitation and other brain-computer interface (BCI) applications. Current methods for analysing the motor characteristics such as acceleration and velocity using magnetoencephalography (MEG) are able to offer satisfying performance, but are facing limitations of the redundancy of channels and limited considerations of causality, which hinder their adaptation to the limited-channel OPM-MEG systems and the methods’ interpretability. Here we propose an algorithm with a causality-aware hybrid decoding network and a voting channel selection method for a 3-D finger movement prediction task. The channel selection employs correlation analysis based on the sliding window average power during the trial period, and bases its selection with reference channels heavily related to the visual and motor area of the brain. The network adopts unidirectional long short-term memory (LSTM) layers which respect the signals’ causality, and convolutional layers to further extract information from the signals. The strategy is able to achieve an average correlation of 0.849±0.050 for the worst axis of the trajectories in an open-source SQUID-MEG dataset with only 18 channels adopted, outperforming current methods in the correlation and channel usages. The algorithm is robust and lighter to prepare in the aspect of both placing MEG channels and data analysis, building a solid foundation for the usage in both traditional SQUID-MEG systems and agile OPM-MEG BCI systems with the potential for real time processing.