Decoding visual object recognition from EEG signals

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Scalp electroencephalography (EEG) signals are crucial for brain-computer interfaces (BCIs) but are often mixed and lack frequency specificity.
• Developing compact and interpretable decoders is essential for clinical EEG applications and efficient object recognition.

Purpose of the Study:

• To identify cortical regions and features that carry discriminative visual information for object recognition.
• To develop efficient, anatomically grounded decoders for BCIs using source-space analysis.
• To compare the efficiency of different feature families and anatomical resolutions for EEG decoding.

Main Methods:

• A source-space decoding pipeline was developed, projecting EEG sensor signals onto anatomically defined cortical regions.
• Four feature families (band-limited power, line length, temporal morphology, couplings) were extracted from regions of interest (ROIs).
• Random Forest classifiers were trained per participant, and decoder generality was assessed across participants.

Main Results:

• A low-dimensional representation based on line length (LL) demonstrated the strongest decoding performance.
• The 24-ROI LL-only model achieved higher accuracy with 92% fewer features than a sensor-space baseline.
• Anatomically informed ROIs improved decoding efficiency and interpretability compared to sensor-space methods.

Conclusions:

• For single-trial, rapid visual presentation decoding, time-domain structure within anatomically defined ROIs captures most discriminative information.
• Line length features provide a compact and effective representation for EEG decoding.
• This neuro-informed approach supports lightweight, interpretable BCIs with clear anatomical attribution.