EEG study pins learning styles to brain wiring, hits 70% accuracy ceiling

A team led by Wiga Maulana Baihaqi (Universitas Gadjah Mada, Indonesia) proposes an EEG-based alternative to traditional learning-style questionnaires, which are slow and subjective. Their paper, accepted at the 4th IEEE International Conference on AI and Mechatronics Systems, compares Phase Locking Value (PLV) functional connectivity against localized power features. Using 28 participants performing Raven's Advanced Progressive Matrices, they trained Support Vector Machines with Leave-One-Subject-Out Cross-Validation plus a 70:30 intra-subject split. For the Visual-Verbal (VV) dimension, subject-independent accuracy reached 70.00%, driven by distinct fronto-occipital polarization. The Active-Reflective (AR) dimension fared worse at 55.56%, suggesting overlapping executive control networks blur group-level boundaries.

The study's most provocative finding is "Systematic Neural Inversion": individual participants with stable connectivity signatures voted up to 20-0 opposite to the global classifier boundary. This means reliable brain-based markers can still contradict a global model. The authors argue rigid one-size-fits-all classifiers are fundamentally bounded by biological diversity. They advocate for adaptive feature transformation techniques—essentially personalized recalibration—before EEG learning-style recognition can generalize robustly across new users. The work also suggests that connectivity patterns, not just regional power, carry meaningful cognitive-style information, even when averaged models struggle.