Random Forest and SHAP analyses identify Gamma and Beta band power at temporal and frontal channels (T3, P3, F7, T4, T4, F8, F7) as the most critical features for EEG familiarity classification¶

Claim · All claims
Evidence · 1 study · associational r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Random Forest importances ranked High Gamma (32.2%), Low Gamma (26.2%), and Beta (15.8%) as top spectral features, with temporal and parietal channels (T3, P3, F7, T4) contributing most; SHAP showed higher Gamma/Beta power at T4, F8, F7 drove predictions toward familiar states. → Isuru Nanayakkara and Thilina Halloluwa 2026

Evidence¶

Isuru Nanayakkara and Thilina Halloluwa 2026¶

Isuru Nanayakkara and Thilina Halloluwa. (2026). Automating Learner Assessment: Benchmarking Machine Learning and Deep Learning Models for EEG-Based Familiarity Prediction. https://arxiv.org/abs/2608.16541 doi:10.48550/arxiv.2608.16541

q2 · i? · associational · r2

Interpretability analysis of the 4-class Random Forest (feature importances) and DNN (SHAP values), both trained on the full 4-class configuration; the article reports Gamma and Beta power at temporal-frontal sites driving familiar-state predictions. No effect size is printed.

"Higher power in the Gamma and Beta bands at T4, F8, and F7 strongly drove the model toward predicting familiar states, neurophysiologically aligning with active memory retrieval processes."

Discussion¶

Related Claims¶

• Trial-level spectral features robustly differentiate stimulus domains (faces vs. equations) but no familiarity-driven spectral shift survives FDR correction — related
• Temporal leakage from shuffled neighboring epochs inflates EEG familiarity classification, with peak weighted F1 dropping from 0.9853 (CNN, stratified) to 0.6038 (CNN, Group K-Fold) — related
• Familiarity recognition is represented by a distributed multivariate spatial-spectral pattern rather than localized single-feature shifts, necessitating multivariate machine learning — related
• Feature-based model predictive power derives from a large and varied set of features rather than a few dominant variables — related
• Domain-separated EEG classifiers decode familiarity within each stimulus category, with peak LOGO Mean Recall of 0.8992 for faces-only (HistGradient Boosting) and 0.8670 for equations-only (CNN) — related