EPECT combines convolution, positional encoding and eigenvalue-guided attention to classify motor-imagery and cognitive tasks from EEG-fNIRS. Across two public datasets under leave-one-subject-out evaluation, reported accuracies ranged from 96.3% to 98.1%, with ablation and integrated-gradients analyses. These were research datasets, not real rehabilitation patients.
Key findings
- Accuracy reached 97.3% for MI, 96.3% for n-back, 98.1% for DSR and 97.9% for word generation. Ablations supported architectural components, while modality-specific attributions were interpreted as aligning with task-relevant regions.
Why this matters globally
If reproducible, the method may reduce inter-subject variability in brain-computer interfaces and support multimodal rehabilitation biomarkers combining electrical and haemodynamic signals.
Thai researcher contribution
A KMUTNB researcher co-developed the architecture with a Korean team, demonstrating Thai participation in AI, engineering and neural-signal processing.
Limitations to consider
Only two public, feature-engineered datasets were used; participant counts and class balance require full-text review. Window dependence can remain even under LOSO. There were no patients, prospective device-shift tests or clinical endpoints, and integrated gradients indicate model sensitivity rather than neural-source proof.
Verify the original sources
MathematicsRead the original article↗DOI: 10.3390/math14132416