Clasificación de emociones en electroencefalograma con Wavelet Scattering y algoritmos de machine learning.
Loading...
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
La clasificación de emociones a partir de señales de electroencefalograma (EEG) es de gran im- portancia para el desarrollo de tecnologías avanzadas en neurociencia afectiva e interfaces cerebro- computadora. Este campo enfrenta desafíos considerables debido a la complejidad y variabilidad de las señales EEG, lo que dificulta la creación de algoritmos robustos y generalizables. En esta memoria ligada al proyecto FONDECYT 13220040 se evaluó el uso de la Transformada Wavelet Scattering (WS) como técnica para la extracción de descriptores de señales EEG, con el objetivo de mejorar el rendimiento de los clasificadores de emociones. El enfoque propuesto busca superar las limitacio- nes de los métodos tradicionales de extracción de características, ofreciendo una mayor capacidad de generalización.
La metodología fue evaluada utilizando la base de datos DEAP, un referente en el estudio de re- conocimiento de emociones con EEG. Los resultados muestran que el clasificador basado en WS, combinado con un Perceptrón Multicapa (MLP) y PCA para la reducción de dimensionalidad, al- canzó un alto rendimiento con precisiones superiores al 95 % en la clasificación de las dimensiones emocionales de Valence, Arousal, Dominance y Liking al utilizar validación cruzada con 5 k-folds. Sin embargo, al aplicar el método Leave-One-Subject-Out (LOSO), se observó una disminución en la precisión, lo que sugiere la necesidad de ajustar la metodología para abordar problemas como la correlación entre canales y la variabilidad intersujeto. En conclusión, la Transformada Wavelet Scat- tering (WS) podría mostrar un potencial significativo en la clasificación de emociones, aunque es crucial mejorar las estrategias de evaluación para asegurar la robustez y generalización del modelo en aplicaciones reales.
The classification of emotions from electroencephalogram (EEG) signals holds great importance for the development of advanced technologies in affective neuroscience and brain-computer interfa- ces. This field faces considerable challenges due to the complexity and variability of EEG signals, making it difficult to create robust and generalizable algorithms. In this thesis, associated with the FONDECYT project 13220040, the use of the Wavelet Scattering Transform (WS) was evaluated as a technique for extracting descriptors from EEG signals. This was done with the objective of improving the performance of emotion classifiers. The proposed approach aims to overcome the limitations of traditional feature extraction methods, offering greater generalization capacity. The methodology was evaluated using the DEAP database, a benchmark in the study of emotion recognition using EEG. The results show that the WS-based classifier, combined with a Multilayer Perceptron (MLP) and PCA for dimensionality reduction, achieved high performance with accuracies exceeding 95 % in the classification of the emotional dimensions of Valence, Arousal, Dominance, and Liking using 5-fold cross-validation. However, when applying the Leave-One-Subject-Out (LO- SO) method, a decrease in accuracy was observed, suggesting the need to adjust the methodology to address issues such as channel correlation and inter-subject variability. In conclusion, the WS shows significant potential in emotion classification, although it is crucial to improve evaluation strategies to ensure the robustness and generalization of the model in real-world applications.
The classification of emotions from electroencephalogram (EEG) signals holds great importance for the development of advanced technologies in affective neuroscience and brain-computer interfa- ces. This field faces considerable challenges due to the complexity and variability of EEG signals, making it difficult to create robust and generalizable algorithms. In this thesis, associated with the FONDECYT project 13220040, the use of the Wavelet Scattering Transform (WS) was evaluated as a technique for extracting descriptors from EEG signals. This was done with the objective of improving the performance of emotion classifiers. The proposed approach aims to overcome the limitations of traditional feature extraction methods, offering greater generalization capacity. The methodology was evaluated using the DEAP database, a benchmark in the study of emotion recognition using EEG. The results show that the WS-based classifier, combined with a Multilayer Perceptron (MLP) and PCA for dimensionality reduction, achieved high performance with accuracies exceeding 95 % in the classification of the emotional dimensions of Valence, Arousal, Dominance, and Liking using 5-fold cross-validation. However, when applying the Leave-One-Subject-Out (LO- SO) method, a decrease in accuracy was observed, suggesting the need to adjust the methodology to address issues such as channel correlation and inter-subject variability. In conclusion, the WS shows significant potential in emotion classification, although it is crucial to improve evaluation strategies to ensure the robustness and generalization of the model in real-world applications.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Biomédico.
Keywords
Electroencefalografía, Aprendizaje de máquina, Emociones