Channel Optimization for EEG-Based Emotion Recognition

Abstract

With the increasing use of human-computer interaction, the need for systems that perceive and interpret human emotions has grown. In this context, electroencephalography (EEG)-based brain-computer interface (BCI) systems are becoming important for objective emotion recognition. However, the multi-channel and complex nature of EEG signals leads to high computational costs and delays during model training, making these systems difficult t o u se in r eal-time applications. This study investigates channel optimization for reducing data dimensionality and computational complexity in multi-channel EEG data. The DREAMER dataset, developed using a 14channel Emotiv Epoc+ device, was used in this study. The raw signals, after being decomposed into sub-bands, were first utilized to estimate the power spectral density (PSD), and then, differential entropy (DE) features were extracted. Then, the channels providing a high contribution to emotion recognition were identified u sing a R andom F orest-based f eature selection model. Finally, various classification m odels w ere t rained on both the 14-channel EEG dataset and the reduced 4-channel (F4, T7, P7, and AF4) dataset. The findings d emonstrate that while reducing the number of channels led to limited losses in accuracy (average 5-6%), depending on the model, it also increased classification speed (Acceleration Rate (AR)) by up to 72.2%.