Um pipeline de aprendizado de máquina para reconhecimento de emoções com base em informações temporais e espaciais de mapas topográficos gerados a partir de sinais de eletroencefalograma

Abstract

Emotion recognition is an area that is becoming increasingly relevant due to its direct implication in society’s daily life. The field seeks better ways to understand how subjective emotions influence human behavior. Given that emotions can manifest non-verbally, brain activity analysis through electroencephalogram signals becomes a relevant approach for emotion identification. The electroencephalogram records the electrical signals generated by the brain cortex through electrodes placed on the scalp. These signals exhibit complex characteristics in spatial, temporal, and frequency domains. Therefore, a prediction pipeline emerges as an alternative for emotion recognition. According to the literature, few studies focus on topographic map images derived from brain signals, especially with information mapped to a structured format. In this context, this work proposes the analysis and development of machine learning models for emotion recognition based on brain topographic maps generated from electroencephalogram signals using the EEGLAB tool. The case study refers to the DEAP, a widely recognized dataset in the area, constructed through electroencephalogram experiments where subjects were exposed to musical and visual stimuli, providing self-assessments on emotional dimensions such as arousal, valence, dominance, and preference. In addition to implementing the pipeline, the main contributions of this study include the presentation of a structured dataset — created based on the mapping of information obtained from topographic map images, containing brain intensity data per region using temporal windows for samples in three frequency bands: Alpha, Beta, and Gama. Finally, the study introduces predictive models for emotion recognition, achieving notable accuracies of 85.46%, 85.05%, 85.92%, and 84.49% for dimensions of arousal, valence, dominance, and preference, respectively. These results were obtained through the K-Nearest Neighbors algorithm using data from all electrodes.