Page 124 - AN-4-2

Advanced Neurology ML for EEG signal recognition

A B

Figure 3. Confusion matrix of CNN performance (%). (A) The performance of the model on the validation data (20%) which the model has previously
encountered. (B) The performance of the model on the test data (10%) which the model has not previously encountered.
Abbreviation: CNN: Convolutional neural network.

A B

Figure 4. Confusion matrix of LightGBM performance (%). (A) The performance of the model on the validation data (20%) which the model has previously
encountered. (B) The performance of the model on the test data (10%) which the model has not previously encountered.
Abbreviation: LightGBM: Light gradient boosting machine.

FPR (1-specificity) at various classification thresholds. The and validation datasets are shown in Figure 5. The CNN
ROC curve provides a comprehensive view of a model’s model achieved an AUC of 0.86 on the testing set and 0.87
performance across all possible thresholds, making it a on the validation set, while the DNN model achieved an
valuable tool for evaluating binary classification tasks, such AUC of 0.84 on the testing set and 0.82 on the validation
as distinguishing between epileptic and non-epileptic EEG set. The interpretations of these results are given in the
signals. The AUC quantifies the overall performance, with following.
values closer to 1 indicating better discrimination. The small differences in AUC values between the
In the context of epileptic seizure detection, ROC validation and testing datasets for both models indicate
curves are particularly important because they allow that the models are robust and generalize well to unseen
clinicians and researchers to assess the trade-off between data. This is critical in clinical applications, as it suggests
sensitivity (correctly identifying seizures) and specificity that the models can reliably classify EEG signals from new
(avoiding false alarms). A high sensitivity is crucial in patients. Robust models ensure consistent performance
clinical applications to ensure that seizures are not missed across different datasets, which is essential for clinical
while maintaining a low FPR is equally important to avoid deployment. A model that generalizes well can be trusted
unnecessary interventions or misdiagnoses. The ROC to assist neurologists in diagnosing epilepsy, reducing the
curves for the CNN and DNN models on both the testing burden of manual EEG analysis and improving diagnostic

Volume 4 Issue 2 (2025) 118 doi: 10.36922/an.7941