tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

From Handcrafted to Deep Representations: ReliefF and DANN Feature Fusion for EEG Emotion Classification

Authors :

Zahra Mahdinezhad¹ Raheleh Davoodi²

1- دانشگاه شهید بهشتی تهران 2- دانشگاه شهید بهشتی تهران

Keywords :

EEG،DANN،ECSMP،emotion recognition،machine learning

Abstract :

Emotion recognition from EEG signals is challenged by noise, variability, and high dimensionality. This study introduces a hybrid framework that combines handcrafted feature optimization with deep representation learning through a Domain-Adversarial Neural Network (DANN). Using the ECSMP database, EEG signals were preprocessed and a diverse set of temporal, spectral, time–frequency, and nonlinear features were extracted. ReliefF reduced these to 43 discriminative indices, which were benchmarked against DANN-derived latent representations across multiple classifiers, including Random Forest, SVM, MLP, and gradient boosting models. Results show that DANN features substantially improved balance and reduced class bias, achieving up to 97.3% accuracy with MLP, CatBoost, and LightGBM. Interpretability was addressed using SHAP, which highlighted the importance of wavelet- and amplitude-based measures (e.g., wavelet length, MAV, Willison amplitude). Unlike prior ECSMP studies requiring extensive multimodal features, our approach achieves competitive six-class classification using only seven EEG channels and a few number of features, highlighting the value of combining optimized handcrafted indices, DANN-based deep representations, and explainable AI for lightweight and interpretable affective computing.