Dataset description
The dataset consists of EEG brainwave data collected from two individuals (male and female) using a Muse EEG headband, which captured signals from four positions. The following muscle activations were detected using dry electrodes: TP9, AF7, AF8, and TP10. The recordings included three valence categories: positive, neutral, and negative for each of the 3-min and 6-min segments of neutral rest data. Figure 2 shows a visualization of the EEG brainwave data, which may correspond to signal changes at regular time intervals. These vertical spikes are proportional to the change in the amplitude of the EEG signal relative to the electrical activity of the brain, depending on the electrode. The x-axis generally refers to time, and the y-axis refers to the amplitude or intensity of the captured EEG signals. Three of these parameters describe the real-time EEG data of rapid changes in brain activity in positive, neutral, or negative states53,54,55.
Sample features fft_0_b through fft_749_b.
Results
In this section, we present classification results obtained from implementing hybrid CNN-LSTM model, CNN-LSTM with random forest, Hybrid CNN-LSTM with fuzzy c-means and hybrid support vector machine (SVM) with fuzzy c-means models using different kernels. The activation functions available are the linear, polynomial, Gaussian, and sigmoid activation functions. Moreover, each kernel provides its own approach in the definition of decision boundaries, which in turn influences the ability of the constructed emotional interface model to differentiate between three emotional states: positive, neutral, and negative.
Experiemnt 1: hybrid CNN-LSTM model
A hybrid model uses Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks to identify emotions from EEG signals. The EEG dataset first undergoes loading before feature extraction proceeds toward reshaping the information to create a 4D array suitable for CNN operation. The data requires label encoding combined with one-hot encoding before it can be used for categorically classifying emotions (positive, negative, and neutral). The spatial features within EEG data are captured through CNN layers followed by max-pooling layers that lower dimensions while maintaining essential information. Time-dependent emotional recognition needs the LSTM layer to process the data array output by CNN layers. The LSTM layer requires the array to be reshaped first. Following the LSTM component, the model adds fully connected layers to extract additional features, and the output layer applies SoftMax activation for predicting emotional classes.
We use categorical cross-entropy loss to train the processed data while optimizing it with Adam. The test set data undergoes performance evaluation for calculating accuracy alongside precision, recall, F1-score, specificity, and drawing the ROC curve. Visualizations of the model performance present three elements, including a confusion matrix with a classification report and ROC curve plots for thorough assessment. The model demonstrates 85% success in test phase emotion detection from EEG signals after obtaining proper training. The combination of CNN and LSTM frameworks enables optimal spatial and temporal pattern discovery, which makes this approach a favorable method to detect emotions using EEG signals.
Table 9 shows the hybrid CNN-LSTM model demonstrates its performance for classifying three categories, namely NEGATIVE, NEUTRAL, and POSITIVE, through evaluation metrics that include precision, recall, and F1-score. The model demonstrates exceptional performance on the negative class since it reaches a precision of 0.81 with a recall of 0.97 and an F1 score of 0.88, thus indicating superior negative instance detection with high recall. The NEUTRAL class shows robust performance, getting an F1 score of 0.84 at the same time it meets 0.79 precision and 0.90 recall scores, which indicates accurate detection of neutral examples yet with limited precision. The POSITIVE class achieves the least effective results among all classes because the model has 0.78 precision yet a very low 0.51 recall, leading to an F1 score of 0.62. The model demonstrates difficulties with recognizing positive examples since it tends to classify them as other classification categories. The combined precision-recall score across all classes reached a rate of 0.79 alongside 0.80 in recall and 0.78 in F1-score, which suggests a decent model execution but needs further improvement, mainly to recognize positive cases accurately. Performance improvement for less frequent classes can be achieved through data balancing methods along with fine-tuning procedures and class-specific weighting strategies.
Figure 3 depicts the accuracy and loss of the suggested CNN-LSTM of the training and validation performance across the epochs. Examining the left plot, it can be observed that the accuracy of training and validation increases steadily, whereby the validation accuracy is always at par or just slightly above training accuracy, which is a good sign of generalization and overfitting with no or minimal overfitting. In the right plot, the training and the validation loss have shown a sharp decline in the first few epochs, and subsequently they have tapered off, and there is no notable difference between the training and the validation loss. This conclusion implies that the model is convergent and does not sacrifice one aspect of learning and generalization over the other. In general, the figure proves the effectiveness and stability of the CNN-LSTM model in identifying emotional states in EEG.
Accuracy and loss for CNN-LSTM model.
Experiment 2: CNN-LSTM with Random Forest
The model uses deep learning algorithms and machine learning methods to execute effective emotion classification of EEG brain signals. The EEG data needs preprocessing before it can fit into a CNN-LSTM architecture through feature reshaping and label encoding procedures. The CNN layers extract spatial features that exist in EEG signals as the LSTM layer identifies temporal dependencies present in these signals. The Random Forest classifier receives extracted features for performing classification after the training process. A training process using a three-class emotional dataset with positive, neutral, and negative classification evaluates with accuracy, precision, recall, and F1-score assessment metrics.
The optimization process combines Principal Component Analysis (PCA) to extract refined features from EEG signals, which reduces their dimensions without losing important details. A Random Forest classifier containing 100 decision trees undertakes the learning process of detecting different emotional states. A high level of model accuracy becomes evident by analyzing the confusion matrix and classification report that shows the results. The fusion of deep learning feature extraction with machine learning classification functions as a robust method for EEG-based emotion detection through an efficient equation between predictive efficiency and computation speed.
Table 10 shows the hybrid CNN-LSTM with Random Forest obtained high success rates for determining NEGATIVE and NEUTRAL samples but demonstrated lower success on POSITIVE classes. The model demonstrates its best results in NEGATIVE class classification at a precision of 0.81, recall of 0.90, and F1 score of 0.85, which ensures high negative instance recognition. The NEUTRAL class demonstrates balanced performance across all metrics because it achieves 0.82 precision, F1 score, and recall metrics, which indicate precisely regulated classification accuracy. The model exhibits the lowest performance among the classes for the POSITIVE category since it obtains a precision of 0.68 and a recall of 0.60 with an F1 score of 0.63. This evidenced the challenge of correctly evaluating positive data points. The combination of CNN-LSTM with Random Forest achieved an overall balanced performance (0.77 Precision, 0.78 Recall, 0.77 F1-score), though the Random Forest addition succeeded in improving NEGATIVE class recall but did not significantly advance POSITIVE class identification. Performance enhancement demands class-specific weighting and data enhancement techniques as well as hyperparameter modification to optimize results, particularly for minority classes.
The training and validation performance of the CNN-LSTM model with the combination of a random forest classifier can be seen in Fig. 4. Training and validation accuracy curves with epochs are exhibited in the left subplot. The two curves show a steady monotonic increase, meaning that the model is learning and generalizing well on unseen data. According to several epochs, the validation accuracy is a little bit higher than the training accuracy, which indicates a good generalization and no overfitting. The appropriate loss curves are depicted by means of the right subplot. The loss of training converges steadily, and the loss of validation converges in the early epochs, but it starts to grow slightly in later epochs, showing some signs of mild overfitting or variance caused by ensemble decision boundaries of Random Forest. In general, the model has a good capacity to classify, and the accuracy and loss on the training and the validation sets are also good, which signifies the effectiveness of temporal-spatial features of deep and enhanced CNN-LSTM and robust decision of Random Forest.
Accuracy and loss for CNN-LSTM with Random Forest.
Experiment 3: hybrid CNN-LSTM model with fuzzy C-means
A combination method applies Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks together with Fuzzy C-Means clustering during data preprocessing when classifying emotions through EEG signals. Data preprocessing begins with the model through loading and preparing the EEG information, during which it utilizes label encoding and one-hot encoding on the target variable. Fuzzy C-Means clustering detects similarities between EEG features to establish three clusters that correspond to the emotions of positive, neutral, and negative. The clustering process enhances the precision level of data handling by allowing greater flexibility for underlying variations while improving classification performance. The data division into training and test groups diagnoses the CNN-LSTM hybrid architecture that identifies spatial characteristics and temporal patterns within the EEG signal waveforms for feature compatibility.
This model contains spatial-feature-extracting CNN layers that precede temporal-dependency-detecting LSTM layers built to handle inherent EEG signal patterns. The neural network contains two convolutional components together with max-pooling and dropout layers to prevent overfitting during training. The data undergoes reshaping after the CNN layers to enable passage into an LSTM layer, which processes the temporal sequences. Equipped with fully connected layers, the model reaches its output point where the emotional categories are classified through a SoftMax output layer. Performance evaluation of the model happens through accuracy, confusion matrix, and classification report metrics, while training occurs with categorical cross-entropy loss and Adam optimizer. The model achieves enhanced performance by adding a Support Vector Machine (SVM) classifier to the feature extraction processes of CNN-LSTM, which enhances the predictive abilities. The end-to-end process for the combined model consisting of training and evaluation followed by visualization generates highly accurate results for EEG signal emotion detection.
According to the results presented in Table 11, the hybrid CNN-LSTM with the fuzzy logic model successfully distinguishes NEGATIVE and NEUTRAL classes but possesses difficulties with POSITIVE classification accuracy. Most negative instances are precisely classified by the NEGATIVE class with the highest achievable designation values consisting of 0.81 precision alongside 0.94 recall and 0.87 F1-score. The NEUTRAL class demonstrates successful classification through model accuracy indicators that reach 0.73 precision and 0.89 recall, together with an F1 score of 0.80. The POSITIVE class demonstrates the most challenging performance metrics since it produces 0.72 precision, 0.43 recall, and 0.53 F1-score, respectively. This difficulty suggests incorrect positive sample detection frequently labels data as different classes. This model’s average performance measures show 0.75 precision, 0.76 recall, and 0.74 F1-score, thus demonstrating that fuzzy logic improves negative and neutral instance detection while facing challenges with positive class recognition. The general classification performance can probably be improved through better adjustments of fuzzy membership functions or utilization of class balancing approaches.
Figure 5 shows the training and validation performance of the hybrid CNN-LSTM model with Fuzzy C-Means clustering. The left subplot depicts the gradual and consistent rise in both accuracy in training and accuracy in validation, where the accuracy of the validation was a bit better than that of training in most epochs. This implies that the model should be generalizing appropriately to unobserved data, and it is not overfitting. The training and validation loss graphs on the right side indicate that the results are good; the curves sharply drop at the first stages of training and smooth out to a minimal point, which means that training is effective, and the process of optimization is stable. The fact that validation loss is lower and more stable than the training loss indicates that before being passed to the classification step, the quality of features was improved by the fuzzy C-means clustering, which leads to better differentiation between the classes of emotions. In sum, this number shows the strength and effectiveness of the hybrid technique of learning discriminative EEG patterns to recognize the emotional states.
Accuracy and loss for hybrid CNN-LSTM model with fuzzy C-means.
Experiment 4: linear SVM with fuzzy C-means
Table 12, which shows the classification results of SVM with a linear kernel, indicates that the model has very high accuracy for all three emotional classes. Again, the precision for the negative and positive classes of events is 0.97, and for the neutral class, it is 0.99, which is almost perfect and represents a very low false positive rate. The same is true for the recall values, which are very high: negative (0.98), neutral (0.99), and positive (0.96), i.e., great recall for the actual instances of each class. For the negative and positive classes, the F1 score was 0.97, while for the neutral class, it was 0.99. The model is very good in terms of both precision and recall. However, the average precision, recall, and F1 score are 0.98, and the average test accuracy is 97.66%, which indicates that an SVM with a linear kernel is well suited for distinguishing between the three emotional states with a negligible error rate.
Experiment 5: polynomial SVM with fuzzy C-means
Table 13, which depicts the class distribution of the selected SVMs with a polynomial kernel, indicates that the classification performance varies between the three emotional classes. The precision, recall, and F1-score of the neutral class were 0.97, 0.96, and 0.97, respectively, which is evidence that the model does equally well in recognizing this emotional state. The negative class was observed to have a slightly lower recall score of 0.83 for this property; this may mean that there are some instances where the class is negative, but the model fails to put them right under this category. On the positive side, the high precision (0.95) indicates that the tendency of the model to correctly classify an instance that is of the positive class is (0.81); however, a recall of (0.93) was achieved, meaning that there are few instances of false positives, but all positive cases are recovered. All three metrics are quite low, with average values of 0.91 for all the classes and an average test accuracy of 90.87%, indicating that the polynomial kernel is reasonably good but not as perfect as the linear kernel in previous outcomes. The relative performance within classes is shown by the polynomial kernel, where increases in accuracy are met with some decline in both precision and recall of classes by the classifier.
Experiment 6: Gaussian SVM with fuzzy C-means
Table 14 shows that the SVM classifier with a Gaussian kernel can classify the emotional data with high accuracy for each of the emotional classes. For the negative class, the model achieves 95% precision in distinguishing between negative instances and 5% precision in distinguishing false alarms out of all nonpositive instances. The recall of 098 reveals that the model is indeed extremely precise at identifying almost all actual negative cases in the data; hence, an F1-score of 096 stably pluses the strength of the developed model in terms of precision without sacrificing the recall of negative cases too much, that is, 098. For the Neutral class, the precision is outstanding at 0.99, so it is almost perfect at identifying objects in the neutral class. The recall of the model is 0.97; it can categorize nearly all true neutral data points. It has an F1-score of 0.98. In the Positive class, the performances for precision, recall and F1-score are 0.94 and 0.93, respectively, indicating that the model performs equally well in classifying positive emotion states. Therefore, on average, the precision, recall and F1-score are 0.96, and the average test accuracy is 95.78%, emphasizing the robustness of the model with all classes of emotions in the dataset.
Experiment 7: sigmoid SVM with fuzzy C-means
Table 15 describes the classification results for the SVM with the sigmoid kernel. For the negative class, the resulting model yields a precision of 0.81, a recall of 0.92, and an F1 score of 0.86, revealing that although the model captures most of the negative samples, the precision decreases slightly. For the Neutral class, the performance is almost flawless, with a precision of approximately 0.98 and the same recall, leading to an impressive F1-score of 0.98, showing that the model developed in this paper provides excellent predictions in identifying the absence of any specific emotion for the individual. The precision of the class positive is 0.89, whereas the recall is 0.76, which gives an F1-score of 0.82. This indicates that the model appears to forget all the positive samples slightly but is good at it when it does so. On average, the precision is 0.89, as are the recall and F1-score, suggesting that the average test accuracy is 88.99%, while the classifier is performing reasonably well and does not optimize performance across all classes with weakness in the positive emotional state.
Discussion
The Linear SVM with Fuzzy C-Means (FCM) model performed best according to Table 16 because it reached 0.98 precision, recall, and F1 score along with 97.66% accuracy, which proved superior to other models (p = 0.0003). The CNN + LSTM + FCM hybrid model achieves 0.75 precision, 0.76 recall, and 0.74 F1-score, but its accuracy of 76% indicates there is potential for improvement (p = 0.0152). When FCM is removed from the CNN + LSTM model, its performance increases slightly to achieve 0.79 precision, 0.80 recall, 0.78 F1-score, and 80% accuracy (p = 0.0098). The performance of the LSTM model acting independently produces the lowest results, including 0.63 precision and recall alongside an F1 score together with an accuracy level of 64%, as indicated by statistical analysis (**p = 0.0325**). When applied as a hybrid model with CNN, LSTM, and Random Forest, the ensemble method delivers balanced performance through 0.77 precision, 0.77 recall, and 0.77 F1-score combined with 78% accuracy (p = 0.0126), indicating ensemble techniques produce effective outcomes. ANOVA tests confirm Linear SVM + FCM provides superior performance to other models, while LSTM achieves the most inferior results. Fuzzy clustering and ensemble methods add value to model accuracy, but further optimizations probably could enhance results even more.
Multiple classification models apply to Fig. 6, which displays confusion matrices that demonstrate their success in predicting class labels. The hybrid CNN-LSTM model maintains a balanced approach toward class identification yet displays several misdiagnosis points when processing particular cases. The hybrid CNN-LSTM with the random forest model reaches enhanced classification precision for selected categories, although it maintains specific areas where misclassification occurs. The Hybrid CNN-LSTM with Fuzzy C-Means (FCM) model presents enhanced classification accuracy because it improves the ability to differentiate between classes and reduces overall misclassification errors. The linear SVM with the FCM model becomes the top-performing model among SVM-based methods since it demonstrates robust correct classification together with minimal misclassification errors through its clear diagonal dominance pattern. When applied with FCM, the polynomial SVM delivers moderate results, yet the Gaussian SVM with FCM generates a well-organized diagonal distribution for accurate prediction. Among the examined model sets, the sigmoid SVM with FCM demonstrates the most classification errors as its off-diagonal values reach a maximum point. Tests indicate that the linear SVM with FCM demonstrates optimal results, which produce both superior precision rates and recall values across all classes through balanced performance metrics.
Confusion matrices for all models.
Figure 7 displays ROC curves for different models through the TPR and FPR axes to demonstrate classification accuracy. The hybrid CNN-LSTM model delivers outstanding discrimination abilities as indicated by its AUC values of 0.95 for Class 0 and 0.93 for Class 1 while measuring a slightly lower 0.77 AUC for Class 2. FCM clustering improves CNN-LSTM model performance to a small degree when implemented. When combined with FCM, the Polynomial Linear Gaussian and Sigmoid SVM models exhibit nearly flawless classification outcomes, resulting in Linear and Gaussian SVM having an AUC value of 1.00. The superiority of SVM models combined with FCM becomes evident as they succeed in achieving better classification results than CNN-LSTM approaches show.
ROC-curve for all models.
The performances of emotion recognition techniques on different datasets of EEG data are compared in Table 17. The ensemble-optimal average empirical mode analysis proposed by Samal and Hashemi achieved an accuracy of 78% on the same dataset, while the authors of Qi Li et al. successfully achieved 95.89% on the DEAP dataset by cross-modal access. While DEAP was classified with an accuracy of 80.83% using direct channel selection by H. Xu et al., Turker et al. achieved an accuracy of nearly 99% using weightless majority voting, Tetromino, mRMR, and both methods. Using the TQWT feature extraction method, Mei-yu et al. achieved 95.33% accuracy on the SEED dataset, while Fernandes et al. achieved 89.97% accuracy using GCNN. Using the same dataset collected by the authors, Lim and Teo achieved an overall accuracy of 90% using both SVM and association rules. However, compared to other methods, the proposed strategy, which combines fuzzy C-means clustering (FCM) with support vector machines (SVMs), yields a high accuracy of 97.66% on the aggregated dataset.
In this paper, a new hybrid architecture of emotion recognition using the EEG signal employing fuzzy C-means clustering and support vector machine (SVM) as a classifier is considered with a variety of kernel functions being used. The fundamental observations include the fact that the linear SVM kernel had the best classification rate of 97.66%, followed by the Gaussian kernel rate of 95.78%, which was better than the classical nonlinear kernels such as the polynomial and sigmoid. Such accuracy using a small sample size of two subjects and three emotion classes proves the competency and generalization power of the proposed approach. This is the novelty of the approach since unsupervised fuzzy clustering and supervised SVM learning can be combined to facilitate the addressing of ambiguous or overlapping emotional states in EEG data. Compared to the conventional approaches of implementing SVM or clustering as isolated datasets, the hybrid architecture can aid in making the decision boundary much more visible and the models simplified. Furthermore, those findings support the clinical importance of the model, and it may be used in brain-computer interfaces (BCIs), mental health assessment tools, and emotion-aware computing systems. Another area of contribution that the study makes to the industry is that it contrasts its performance with more than 25 new models that have been released in 2024–2025 and takes an in-depth examination of computational efficiency, practical applicability, and potential real-time deployment.
In alignment with our initial hypothesis, our results provided strong empirical support. The linear kernel SVM had reached an accuracy of 97.66%, confirming the postulation that a simpler decision boundary with the help of fuzzy-clustered preprocessing can give high accuracy in classification. The Gaussian kernel also performed well (95.78%), meaning that smoothness in nonlinear separability can even work with some data distributions. Such results support our hypothesis that the combination of fuzzy logic and classical ML techniques is the most effective method of detecting emotions because it is able to identify the overlapping emotional conditions in the EEG signals. Nevertheless, decreasing performance shown with sigmoid and polynomial kernels indicates that too complex boundaries are not easily generalizable on small or difficult-to-learn EEG data. This critical view validates the power of our hybrid paradigm as well as documenting its shortcomings and the requirement of the contextual selection of kernel.
