4.9. Results and Evaluation

4.9.1. Classification Performance

The evaluation of classifiers was implemented using Group K-Fold cross-validaion with 10 folds, ensuring good testing performance by keeping participant data consistent with each fold. The table 3.1 below summarizes the mean accuracy scores for each classifiers tested.

Table 1: Performance Comparison of Classifiers for EEG-ECG Stress Detection

The evaluation of classifier performance in EEG-ECG stress detection shows variation in accuracy and runtime across different models. High-performing classifiers, such as RBF SVM, AdaBoost, Decision Tree, and Nearest Neighbors, achieved mean scores above 0.95, showing strong potentials in diffentiating stress from non-stress states.

Notably, RBF SVM and AdaBoost both reached the highest accuracy (0.9833), indicating their effectiveness in capturing the complex patterns in the data. However, AdaBoost's runtime (0.0040s) was higher than other models, likely due to its ensemble nature that combines multiple weak learners to enhance classification accuracy. In contrast, Decision Tree achieved similar high accuracy with a much faster runtime (0.0004s), suggesting that simpler models can be both efficient and effective when the data structure aligns with the model’s strengths.

Moderate and low-performing classifiers, such as Random Forest, Gaussian Process, Linear SVM, and Naive Bayes, demonstrate some limitations in handling EEG-ECG data for stress classification. Linear SVM and Naive Bayes, both with mean scores of 0.6667, seem to struggle with the non-linear relationships within the data, resulting in lower performance.

The Gaussian Process classifier, despite its more complex structure, achieved a moderate accuracy of 0.8167, possibly limited by the high dimensionality of the features or the relatively small dataset. Random Forest also performed moderately, with a score of 0.8667, which may reflect constraints related to ensemble depth or feature diversity.

Neural Net, with a mean score of 0.9333, balanced high accuracy with efficient runtime.

4.9.2. Challenges and Observations

The main challenges in this project arose during the implementation of data preprocessing and feature extraction steps, rather than the model training itself. Handling and preparing the EEG and ECG signals required significant effort to ensure data quality and proper format for model input. The loading and parsing of data from the DREAMER dataset involved identifying the correct structure for each signal type, which proved complex due to variations in format and the hierarchical nature of the dataset. This aspect required careful attention to avoid data misalignment or incorrect structuring, which could impact downstream analysis.

The preprocessing for ECG data was particularly time-intensive. Steps such as signal cleaning, peak detection, and the extraction of heart rate variability (HRV) metrics demanded substantial processing time. While these steps were crucial to capture relevant physiological indicators for stress classification, the high processing time highlights a key limitation: real-time application may not be feasible with current methods. This limitation points to a need for optimization in ECG preprocessing or exploring lightweight algorithms if the aim is to develop a system for real-time stress detection.

4.9.3. Improvement Suggestions for Stress Detection

To enhance the accuracy and efficiency of stress detection, few improvement strategies could be implemented. First, optimizing the preprocessing and feature extraction pipelines, particularly for ECG, could significantly reduce computational time, making real-time applications more feasible. Techniques such as adaptive filtering [134], [135] or faster peak detection algorithms [136] might streamline ECG processing while retaining data quality. Additionally, exploring alternative feature sets that incorporate time-frequency analysis or non-linear metrics may better capture subtle stress-related changes in EEG and ECG signals.