Page 35 - BH-2-2

Brain & Heart Predictive modeling using electroencephalogram

The findings from using all three classifiers together and PCA. Subject 2 appeared to possess the most driving
revealed that, in some cases, the classifiers detect different experience, as indicated by their consistently accurate
driving modes across all scenarios. This is due to the classifications by both KNN and PCA. Moreover, it
fact that classifiers’ performance varies depending on was observed that Scenario 3 was commonly labeled as
the scenario and rarely achieves a perfect score. When ineffective by most individuals using reliable classifiers.
tested on the dataset, both the KNN and PCA classifiers According to the observations, these incidents occurred
demonstrated 100% efficiency, and it was discovered that when the driver was either too close to or too far from a
the driving modes observed in these instances by both slower vehicle. Erratic maneuvering led to the observed
classifiers are, in most cases, the same. As a result, the inexpert mode. Overall, inexpert driving emerged as the
collected master data reflects the common characteristics most categorized style, possibly influenced by participants’
observed across all subjects. There could be a number of lack of familiarity with the virtual city environment in
factors contributing to the fact that aggressive driving is which they are attempting to drive.
the least well-defined subtype of driving. The fact that The results from KNN, PCA, and NN classifiers
all the drivers are aware that their every move is being for the same individual in identical situations exhibit
recorded could explain this. In addition, the driving dissimilarities. The built-in classifiers simply lack
scenarios may not effectively stimulate aggressive driving sufficient accuracy for this purpose. Notably, only KNN
tendencies. achieves a classification success rate of 98.02% on average.
Intriguingly, all classifiers agreed that Subject 1 was Furthermore, KNN once again recognized the least
the least efficient driver, which aligns with their status experienced driver as ineffective, while one of the most
as a novice driver in training. The majority of Subject 1’s experienced drivers was recognized as alert. Notably,
inexpert classifications occurred during turns, revealing all the female drivers maintained a relatively stable (and
a lack of proficiency in this aspect. Conversely, Subject 3 generally upbeat) emotional state throughout the majority
was consistently categorized as perceptive by both KNN of the experiment. Conversely, only drivers fell into the
categories of aggressive behavior, negative valence, and
Table 8. K‑nearest neighbor classifier high arousal. Incorporating explainable AI techniques into
the sleep state classification model could provide valuable
Driving scenario Subject 1 Subject 2 Subject 3 insights into the decision-making process and enhance the
1 Inexpert Inexpert Inexpert interpretability of the results. One approach to achieving
15
2 Inexpert Aggressive Keen this is through the use of feature visualization techniques,
3 Inexpert Inexpert Inexpert such as layer-wise relevance propagation (LRP) or saliency
4 Inexpert Keen Keen maps, which highlight the regions of input data that
5 Inexpert Inexpert Keen contribute most to the model’s predictions. By visualizing
the feature spaces at different layers of the ResNet
architecture, researchers can gain a better understanding
Table 9. Principal component analysis classifier of how the model extracts and processes information from
the multichannel EEG signals.
Driving scenario Subject 1 Subject 2 Subject 3
1 Inexpert Keen Inexpert 5. Conclusion
2 Inexpert Inexpert Keen This research emphasizes the importance of employing an
3 Inexpert Inexpert Inexpert ML model that processes raw EEG signals to accurately
4 Inexpert Inexpert Keen estimate a user’s mental state, especially in the context of
5 Inexpert Keen Keen BCI programs. Our study intends to improve the efficacy
of predictive modeling by employing a methodical
approach and leveraging features such as band powers,
Table 10. Dual‑layer neural network classifier
averages, and root-mean-squared extracted from EEG
Driving scenario Subject 1 Subject 2 Subject 3 signals. Among the classifiers tested on the BCI IV
1 Inexpert Inexpert Inexpert Competition-I dataset, KNN, PCA, and NN performed
2 Inexpert Inexpert Aggressive the best. Notably, KNN achieved an average classification
3 Inexpert Inexpert Inexpert success rate of 98.02%. In addition, our study expands
the application of our methodology to human driving
4 Inexpert Inexpert Keen behavior, showcasing its versatility and promise across
5 Inexpert Keen Keen multiple fields. Further, research into BCI and EEG

Volume 2 Issue 2 (2024) 11 doi: 10.36922/bh.2819