Emotion detection using AI with explainable AI based perception: A comparative analysis between machine learning and deep learning

Background Emotion recognition from electroencephalography (EEG) signals has emerged as an important research area within Affective Computing and Human-Centered Artificial Intelligence due to its potential applications in mental health assessment, adaptive human-computer interaction, and emotion-aware intelligent systems. EEG signals provide a non-invasive and reliable means of capturing neural responses associated with emotional states, while Artificial Intelligence (AI) enables efficient analysis of complex and high-dimensional physiological data. Methods This study investigates EEG-based emotion recognition using two publicly available Kaggle datasets: the Emotion Bandpower Dataset and the Gamma-Theta Enhanced EEG Dataset. Both datasets contain statistical and spectral features extracted from multiple EEG frequency bands and are labelled for binary emotion classification tasks involving Relaxed and Funny/Happy states. A comprehensive comparative analysis was conducted using five Machine Learning (ML) algorithms, namely Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and XGBoost, together with six Deep Learning (DL) architectures including MobileNetV2, ResNet50, VGG16, LSTM, EEG-Former, and Swin Transformer. To improve model transparency and trustworthiness, Explainable Artificial Intelligence (XAI) techniques, including Local Interpretable Model-Agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), and Partial Dependence Plots (PDP), were employed to identify influential EEG features. Furthermore, a quantitative explanation consistency framework based on Spearman Rank Correlation, Feature Overlap Score, and Agreement on Feature Significance was applied to evaluate the reliability of explanations generated by LIME and SHAP. Findings Experimental results demonstrated that Random Forest achieved 99% classification accuracy on both datasets, while XGBoost attained 99% accuracy on the Emotion Bandpower Dataset, outperforming the remaining models. The analysis revealed that Gamma and Delta-related features played a significant role in emotion classification. The findings demonstrate that, despite minor variations in feature rankings, the explainability methods consistently identified a common set of important EEG attributes. Interpretation Overall, the proposed framework provides a transparent, accurate, and reliable approach for EEG-based emotion recognition, supporting the development of trustworthy emotion-aware systems for healthcare, neurotechnology, and human-computer interaction applications.