NeuroEmotionNet: A lightweight and interpretable 1D CNN Framework for Real-Time EEG Emotion Classification

Karim, Mohammad Hasnatul and Pranta, Al Shahriar Uddin Khondakar and Jashim, Farhan Bin and Bhuiyan, Md Imranul Hoque and Masum, Abdullah Al (2025) NeuroEmotionNet: A lightweight and interpretable 1D CNN Framework for Real-Time EEG Emotion Classification. International Journal of Science and Research Archive, 15 (2). pp. 1442-1457. ISSN 2582-8185

Mental health disorders, including anxiety, depression, and emotional dysregulation, affect hundreds of millions globally, yet early diagnosis remains a challenge due to the reliance on subjective assessments such as psychometric tests and clinician observations. While EEG-based emotion recognition offers a non-invasive, cost-effective alternative, existing approaches are limited by small and imbalanced datasets, handcrafted features, and lack of real-time deployability. These gaps hinder the development of scalable and clinically relevant emotion detection systems. To address these limitations, this study proposes a machine learning framework for real-time, accurate emotional state classification using EEG signal analysis. The dataset comprises EEG recordings from 300 participants (158 male, 142 female), labeled across four emotional states: Positive, Neutral, Anxiety, and Depression. Signals were collected using an 8-channel EEG device and decomposed into frequency bands (alpha, beta, gamma) using Discrete Wavelet Transform (DWT), with Shannon Entropy applied for complexity analysis. Data augmentation techniques—Generative Adversarial Networks (GAN), SMOTE, and ADASYN—were used to generate 20,000 synthetic instances per method, addressing class imbalance and data scarcity. A comparative evaluation was conducted across nine classical and deep learning models, including Support Vector Machine, Decision Tree, Random Forest, and a 1D Convolutional Neural Network (NeuroEmotionNet). The models were assessed using accuracy, precision, recall, F1-Score, Matthews Correlation Coefficient (MCC), and PR AUC, with latency tracked for real-time viability. NeuroEmotionNet achieved the highest performance with an F1-Score of 98.16%, MCC of 98.2%, and inference time under 3.2 milliseconds on the combined augmented dataset. The novelty of this study lies in its integration of hybrid feature extraction, multi-strategy augmentation, and real-time deployment. A fully functional web application was developed, making this the only study among comparable works to achieve both high accuracy and practical applicability. This research paves the way for scalable, interpretable, and real-time emotion monitoring systems in mental healthcare environments.