A Multilevel Feature-based Emotion Classification Technique using Electroencephalogram Signals

Kabita Devi (Department of Electronics and Communication Engineering, Tezpur University, Assam, India)
Bhabesh Deka (Department of Electronics and Communication Engineering, Tezpur University, Assam, India)
Debarun Chakraborty (Department of Electronics and Communication Engineering, Tezpur University, Assam, India)

Journal of Electronics, Electromedical Engineering, and Medical Informatics
Vol 8 No 4 (2026): October

Emotion dysfunction can be an indication of mental health conditions like depression, stress, anxiety, etc. Therefore, appropriate recognition of human emotion is crucial for early monitoring of mental well-being. This study aims to develop a robust multilevel feature concatenation framework that effectively integrates handcrafted and deep learning (DL) features to improve electroencephalogram (EEG)-based emotion classification. The proposed framework is evaluated on two publicly available benchmark EEG datasets: DEAP and DREAMER. Initially, handcrafted features extracted from the statistical, time, frequency, and nonlinear domains are categorized into low-level features (LLF) and mid-level features (MLF). To select the most relevant handcrafted features, statistical analysis, minimum Redundancy Maximum Relevance (mRMR) and the Pearson correlation coefficient (PCC) are employed. In parallel, a two-dimensional convolutional neural network (2D-CNN) is used to learn hierarchical deep features, which are grouped into LLF, MLF, and high-level features (HLF). The handcrafted LLF and MLF are then concatenated with the corresponding DL-based LLF, MLF, and HLF to construct a discriminative multilevel feature representation. The resulting feature vector is classified using Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and XGBoost classifiers under a 10-fold and leave-one-subject-out (LOSO) cross-validation protocol. Performance is evaluated using accuracy, precision, recall, F1-score, and balanced accuracy. Experimental results demonstrate that the proposed framework consistently outperforms existing state-of-the-art methods. Among the evaluated classifiers, SVM achieves the highest performance, yielding accuracies of 93.07% and 98.70% for 4-class and 2-class emotion classification, respectively, on the DEAP dataset, and 93.85% and 98.78% for 4-class and 2-class emotion classification, on the DREAMER dataset. These results demonstrate that concatenation of handcrafted with DL-based features at different abstraction level significantly enhances feature discriminability and classification robustness. The proposed framework provides an effective solution for EEG-based emotion recognition, with promising applications in affective computing, human–computer interaction, and intelligent mental health monitoring.