CEP-MI: a novel cue-execution dual-stage motor imagery paradigm with PGCANet decoding architecture

Published October 22, 2026. A new motor-imagery task and EEG decoding model distinguished imagined movements with encouraging accuracy in healthy volunteers. It is an engineering and physiology result, not evidence that the method improves rehabilitation outcomes in patients.

Li J, Yu P, Qiao H, Zheng Z, Wang C, Weng Z, Gong C.
Frontiers in Human Neuroscience · Publication date: 2026-10-22
Added to site: 2026-10-22
DOI: 10.3389/fnhum.2026.1908608

Published October 22, 2026. Added to Neurofeedback.com October 22, 2026.

Modalities

EEG, Brain-computer interface, Motor imagery

Clinical focus

Motor intention decoding, Participant engagement, Neurorehabilitation methods

Population

Healthy young adults performing hand and foot motor-imagery tasks

Sample

97 participants in the new dataset, plus the public nine-participant BCI Competition IV-2a dataset

Intervention / method

A countdown-anchored motor-imagery task was followed by slight motor execution within the same trial; 64-channel EEG was analyzed with ERP and ERD/ERS methods and decoded with the proposed dual-branch PGCANet architecture

Comparator

Classic motor imagery without the execution stage, established decoding algorithms and ablated versions of the proposed model

Clinician summary

PGCANet reached optimal mean accuracy of 85.77% for binary classification and 82.47% for four-class classification. In the new dataset, cue-execution versus classic-task accuracies were 85.77% versus 80.82% for left-right hand imagery and 82.06% versus 80.00% for left-right foot imagery; ERP and ERD/ERS analyses also showed task-related differences.

How to interpret the findings

Participants were healthy, aged 18 to 28, and 80 of 97 were male. The work evaluated offline EEG decoding and neural patterns, not patient rehabilitation, functional outcomes, online device control or durability. The paper notes that only hand and foot imagery were studied and calls for evaluation in other BCI paradigms.

Equipment and disclosures

EEG was recorded with a 64-channel Neuroscan system at 1,000 Hz and processed with EEGLAB; 62 scalp channels were retained. This equipment is reported from the paper and is not attributed to BrainMaster.

Supported by the National Natural Science Foundation of China grant 62502244 and Inner Mongolia Natural Science Foundation grant 2023QN06002. The authors declared no commercial or financial conflicts and reported no generative-AI use in the manuscript. No correction, retraction or expression-of-concern relation appeared in the retained Crossref record; this is a bounded check. AI-assisted review.

Related evidence

Compare findings from the same clinical area, including mixed results.