Transfer Learning in Trajectory Decoding: Sensor or Source Space?

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interface (BCI) systems require extensive calibration, limiting their practical application.
• Transfer learning offers a potential solution to reduce BCI calibration time by leveraging existing data.
• Continuous hand trajectory decoding is a key BCI application area with significant potential.

Purpose of the Study:

• To investigate the effectiveness of across-participant and across-session transfer learning for minimizing BCI calibration time.
• To evaluate different transfer learning strategies for continuous hand trajectory decoding.
• To compare the performance of sensor-space versus source-space features in transfer learning scenarios.

Main Methods:

• Reanalysis of existing BCI data from 10 participants across three sessions.
• Utilized a leave-one-participant-out (LOPO) model as a baseline.
• Employed Recursive Exponentially Weighted Partial Least Squares Regression (REW-PLS) for efficient model training.
• Compared four scenarios: generalized with no update (Gen), generalized with cumulative update (GenC), individual with cumulative update (IndC), and individual non-cumulative update (Ind).

Main Results:

• Generalized models (Gen, GenC) performed below chance level, indicating poor cross-participant and cross-session generalization.
• Individual models (IndC) did not significantly improve performance over non-cumulative models (Ind).
• The best decoding performance was achieved using individual models trained with sensor-space features, outperforming source-space features.

Conclusions:

• Current transfer learning approaches demonstrate an inability to generalize effectively across participants and sessions for this hand trajectory decoding task.
• Sensor-space features provided superior performance compared to source-space features in individual BCI models.
• The decoding patterns in individual models were localized around the precuneus, suggesting task-specific neural representations.