Multi-Task EEG Signal Classification for Emotion Recognition and Epilepsy Detection Using Intelligent Learning Models

Abstract

Electroencephalography (EEG) provides rich information and a representation of brain activity for intelligent healthcare applications. This study proposes a framework for automated emotion recognition and epileptic seizure classification using machine learning (ML) and deep learning (DL) techniques. The framework was tested on three EEG datasets to cover areas of anomaly detection, multi-class emotion recognition, and binary seizure prediction. EEG signals were first normalized and then segmented into fixed-length windows depending on each participant's recording traits, thus maintaining subject-specific temporal patterns. Models’ evaluation was done with properly separated training and testing data to guarantee a trustworthy performance assessment.

For emotion recognition, the model based on Gated Recurrent Units (GRU) obtained 96% accuracy on the test data, whereas ensemble learning with Random Forest got 98%, thereby proving its excellent discriminative power on structured EEG features. Anomaly detection without supervision through Histogram-Based Outlier Score (HBOS) was able to detect the abnormal single-channel EEG segments accurately. In seizure classification, a convolutional neural network (CNN) trained on log, scaled time, frequency spectrograms yielded 95. 75% accuracy with an AUC of 0. 996 on the test dataset, thus successfully differentiating interictal and preictal states. The findings confirm that ML models provide robust and computationally efficient performance on engineered EEG features, whereas DL models effectively capture complex temporal and spectral representations across multiple EEG analysis tasks