Emotion Recognition from Multivariate EEG Feature Signals Using the E-AADEP Framework: Enhanced Adaptive Attention-Based Dream Emotion Predictor

Abstract

This paper presents a comprehensive experimental study of the Enhanced Adaptive Attention-Based Dream Emotion Predictor (E-AADEP) applied to the emotions.xls dataset comprising 2,131 multivariate EEG-derived feature samples across 2,548 feature dimensions with three balanced emotion classes: NEGATIVE (n=707), NEUTRAL (n=716), and POSITIVE (n=708). The E-AADEP framework integrates seven tightly coupled stages: (1) EEG signal acquisition during REM sleep with age-group (A1–A4) and gender metadata; (2) signal preprocessing via band-pass filtering, Independent Component Analysis (ICA) artifact removal, and Z-score normalization; (3) parallel feature extraction combining frequency-domain Power Spectral Density (PSD) features across Delta, Theta, Alpha, and Beta bands with entropy-based non-linear measures (Shannon Entropy, Sample Entropy, Fuzzy Entropy); (4) attention-based adaptive feature weighting using a learned Softmax attention mechanism; (5) demographic adaptation encoding age group and gender into a conditioning vector; (6) Random Forest ensemble classification with T=200 trees using Gini impurity-based node splitting and majority voting; and (7) Softmax output layer producing final class probability estimates. Experimental evaluation on the emotions.xls benchmark yields classification accuracy of 99.30%, weighted F1-score of 99.30%, and AUC-ROC of 0.9998 on the held-out test set (n=427 samples). The NEUTRAL class achieves perfect classification (F1=100%), with near-perfect performance for NEGATIVE (F1=98.95%) and POSITIVE (F1=98.93%). These results demonstrate the effectiveness of the E-AADEP pipeline for automated EEG-based emotion recognition.