Bridging motor execution and motor imagery BCI paradigms: An inter-task transfer learning approach

Abstract

Motor imagery (MI)-based brain–computer interfaces (BCIs) decode movement imagination from brain activity, but improving decoding accuracy from electroencephalography (EEG) remains challenging. MI-based BCIs require calibration runs to train models; however, participant engagement cannot be externally verified. Motor execution (ME) is more straightforward and can be supervised. Deep learning (DL) leverages transfer learning (TL) to bypass calibration. This is the first work to explore wether a ME-trained DL model can reliably classify MI without finetuning to the MI task, thereby achieving direct TL between ME and MI tasks. We employed EEGSym, a DL network for inter-subject TL of EEG decoding, evaluating three scenarios: ME to MI, ME to ME, and MI to MI classification. We analyzed performance correlation between scenarios, and used shapley additive explanations (SHAP) to elucidate model focus patterns learned from ME or MI data. Results show that DL models trained on ME data and tested on MI perform comparably to those trained on MI data. A significant positive correlation was found between performance in ME and MI tasks for models trained on ME data. Explainable artificial intelligence (XAI) techniques revealed robust correlation between patterns in ME and MI tasks. However, between 0.5 to 1 s, the ME-trained model focused on the contralateral central region, while the MI-trained model also targeted the ipsilateral fronto-central region. Our findings demonstrate the viability of inter-task TL between ME and MI using DL models in BCI applications. This supports using ME-trained models for MI tasks to enhance targeted learning of brain activation patterns.