Elektroensefalografi (EEG) sinyali kullanılarak yapay zeka tabanlı duygu kestirimi = Emotion prediction using artificial intelligence based on electroencephalography (EEG) signals

Abstract

Understanding and defining the neurological foundations of emotions is increasingly important in the field of human-machine interaction. EEG (Electroencephalography) signals stand out as a high temporal resolution method for analyzing these emotional states. EEG enables the objective assessment of individuals' emotional and mental states by measuring the brain's electrical activities. However, the inherent complexity of EEG signals and their susceptibility to external factors make accurate and consistent analysis challenging. Therefore, the use of artificial intelligence methods in EEG-based emotion recognition systems has gained significant momentum in recent years. In this study, emotion recognition processes were carried out based on EEG signals using the SEED EEG dataset, and in this context, the Ensemble Bagged Trees model was proposed. The SEED dataset, which forms the basis of this study, contains 62-channel EEG signals from 15 healthy participants. These participants were shown 72 film clips to evoke various emotional responses, and EEG data were recorded at a sampling rate of 1000 Hz during the sessions. In the preprocessing stage, EEG signals were filtered with a bandpass filter of 0.3-50 Hz, downsampled to 200 Hz, and various artifacts (such as eye blinks and muscle movements) were removed. These preprocessing steps enhanced the signal-to-noise ratio, allowing for more reliable and accurate analyses. Specifically, after downsampling, the signals from critically identified six symmetric temporal electrodes (FT7, FT8, T7, T8, TP7, TP8) were analyzed. In the feature extraction stage, EEG data were divided into fundamental frequency bands known as delta (0.5-4 Hz), theta (4-8 Hz), alpha (8-12 Hz), beta (12-30 Hz), and gamma (30-50 Hz). For each band, time domain, frequency domain, and statistical features were calculated, and these features were used to train emotion recognition models. A total of 34 different features were extracted, including parameters such as autocorrelation, total harmonic distortion, kurtosis, skewness, standard deviation, and median. Especially, features such as Crest Factor and Impulse Factor are crucial for understanding the peak values and impulsive characteristics of the signal. The Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) help us analyze the dependencies of the signal over time. Time-domain features express how the signal changes over time, considering the values at specific points and their distribution over time. Frequency-domain features represent the frequency components of the signal, taking into account the energy distribution in the frequency spectrum and the frequency characteristics of the signal. Features that include basic statistical parameters describe the distribution and overall characteristics of the signal. Spearman's correlation coefficient was used to rank the features, and the most significant features were identified. In this process, the impact of each extracted feature on emotion recognition success was evaluated, and features with the highest correlation coefficients were selected. In the classification stage, two different approaches were followed. In the first classification, features extracted from all frequency bands were used to estimate a general emotional state. In this process, 80% (2160 samples) of the 2700 samples were used for training, while 20% (540 samples) were set aside for testing. The overall model accuracy was calculated as 94.4%. This high accuracy indicates that the model can successfully distinguish between the four main emotion classes (neutral, sadness, fear, and happiness) and generally demonstrates reliable performance. In the second classification, each frequency band was analyzed separately to investigate how emotion estimation could be achieved using a single frequency band. In these analyses, the Beta band provided high accuracy rates in the fear (91.40%) and happiness (84.30%) classes. In contrast, the Delta band showed the lowest accuracy rates across all classes, indicating that low-frequency bands might contribute less to emotion detection. For example, the accuracy rates were calculated as 69.9% for the Delta band, 74.8% for the Theta band, 73.8% for the Alpha band, 75.0% for the Beta band, and 73.1% for the Gamma band. Examining the effects of different frequency bands on emotion recognition tasks is a significant dimension of this study. The combined use of all frequency bands generally provided higher accuracy rates, as this method creates a richer data representation by combining the different information provided by each band. For instance, the use of all bands together resulted in an accuracy of 96.94% in detecting the neutral state. Similarly, accuracy rates of 97.78%, 97.13%, and 97.04% were achieved for sadness, fear, and happiness, respectively. These results indicate that the model can classify emotional states quite accurately by using all bands together. When individual bands were considered, the low accuracy rate of 69.9% provided by the Delta band suggests that this band offers limited information, particularly in the analysis of low-frequency signals. The Delta band may reflect slower brain activities associated with deep thinking or resting states, but it has limited impact on distinguishing emotional states. In contrast, the Beta band provided high accuracy rates of 91.40% in fear detection and 84.30% in happiness detection. The Beta band is associated with high-frequency brain activities, indicating that it may be more decisive in detecting emotional responses. The Theta (74.8%), Alpha (73.8%), and Gamma (73.1%) bands provided similar accuracy rates, showing effectiveness in general emotion detection, though no significant superiority was observed among these bands. The findings of the study indicate that EEG-based emotion recognition systems have broad application potential in fields such as human-machine interaction, psychological research, and marketing strategies. These AI-based methods contribute significantly to future research and applications by enabling the accurate detection of emotional states. The high accuracy rates of the Ensemble Bagged Trees model, in particular, support the effective use of such systems in practical applications. This study not only demonstrates that EEG data can be effectively used in emotion recognition and analysis but also highlights how the combination of different frequency bands can enhance model performance. These findings suggest that EEG-based AI methods can be applied more widely and contribute significantly to the literature in this field. In conclusion, the Ensemble Bagged Trees model used in this study has demonstrated overall high accuracy and successful performance in the EEG-based emotion recognition process. However, considering the different effects of specific frequency bands on emotion detection, there are various avenues for further optimization and improvement of these systems. Future studies may focus on testing this model with different datasets and artificial intelligence methods, as well as exploring enhancement methods for bands with lower performance. Each new study in this field will contribute to further developing EEG-based emotion recognition systems and increasing their application potential across various domains.