Optimizing EEG - based Emotion Recognition with a multi-modal ensemble approach

Emotions significantly influence human communication and are intricately associated with cerebral activity, as evidenced by EEG data, which enhances human-machinemachine interactions, especially in Brain–Computer Interfaces (BCI). Considering this potential, attaining high precision in emotion recognition from EEG data continues to pose a considerable barrier, but such progress could yield critical insights for diagnosis of mental and behavioural illnesses and facilitate treatment decision-making in clinical environments. Recent advancements in deep learning techniques have established them as formidable instruments for improving classification efficiency in BCI systems. This study presents a novel hybrid model that integrates Encoder and Gated Recurrent Unit networks to enhance the accuracy of emotion categorisation using EEG data. The Encoder from autoencoder is utilised for dimensionality reduction and feature extraction, generating a compressed latent representation of the input signals. The latent characteristics are subsequently input into the GRU, which captures the temporal dynamics of EEG data to forecast emotional states. The Encoder is used for dimensionality reduction and feature extraction, while GRU is used to capture the temporal dynamics of EEG data. This paper introduces an ensemble technique employing Soft Voting, which integrates predictions from three deep learning models: LSTM network, a novel Deep Neural Network, and the hybrid Autoencoder-GRU model. This ensemble seeks to enhance classification accuracy by utilizing the complimentary capabilities of each model. The method was validated using the EEG Brainwave Dataset. This approach was validated using the EEG Brainwave Dataset, achieving a superior classification accuracy of 97.42%, surpassing individual model performances (LSTM: 97%, GRU: 97%, Autoencoder-GRU: 94%, DNN: 95%). The results confirm that the ensemble model effectively integrates complementary learning patterns from multiple architectures, leading to a more robust and generalizable classification system. These advancements make a significant contribution to EEG-based emotion recognition, facilitating more accurate mental health diagnostics, improved BCI applications, and enhanced human–machine interactions. The findings underscore the potential of ensemble deep learning models in addressing complex classification tasks where precision is critical.