Page 122 - AN-4-2

Advanced Neurology ML for EEG signal recognition

metrics underscore LightGBM’s suitability for accurate and LGBM classifier models used for the classification of
and reliable classification tasks. Because of its superior epileptic and non-epileptic EEG signals, highlighting their
performance, LightGMB was selected for further analysis learning behaviors and generalization capabilities across
to be compared to the CNN and DNN. epochs and training instances.
3.1. Learning curves The performance of the three models – DNN, CNN,
and LGBM classifier – offers important insights into
Learning curves are a fundamental tool used in machine their applicability for distinguishing epileptic from non-
learning (ML) to assess the performance of a model during epileptic EEG signals. Both the DNN and CNN models
training and validation. They provide insights into how exhibit the expected trend where training accuracy
well a model generalizes to unseen data and whether it is steadily increases and plateaus around 98%. Similarly,
overfitting or underfitting. A typical learning curve plots their validation accuracies stabilize at approximately 82%
training and validation accuracy (or loss) against epochs and 80%, respectively, maintaining a reasonable gap from
or the number of training instances. Ideally, the training the training curves. This behavior indicates that both
and validation accuracy should increase steadily and models are learning meaningful patterns from the EEG
plateau, with the validation accuracy consistently lower data without severe overfitting, making them suitable for
than the training accuracy, reflecting the model’s ability to clinical applications where generalization to unseen data
generalize while avoiding overfitting.
is critical. The slight fluctuation in validation accuracy,
In the context of this study, learning curves are particularly in the CNN model, may suggest sensitivity
particularly relevant for evaluating the performance of to hyperparameter selection or inherent variability in
ML models in distinguishing epileptic from non-epileptic the EEG dataset but remains within acceptable limits for
EEG signals – a challenging clinical task requiring high clinical decision-making.
diagnostic accuracy and generalization. Monitoring these The LGBM classifier, however, demonstrates a perfect
curves ensures that the models are learning robust features
from the EEG data rather than memorizing patterns training score of 1.0 across all training instances, which
specific to the training set, which is critical for reliable is a potential sign of overfitting to the training data.
deployment in clinical settings. Figure 1 illustrates the The cross-validation score stabilizes around 85%, with
training and validation performance of the DNN, CNN, minimal improvement as the number of training instances
increases. While the LGBM classifier achieves a slightly
higher validation score, its inability to show a steady rise
and plateau in both training and validation curves raises
concerns about its ability to generalize effectively to new
EEG data. This behavior may stem from insufficient
regularization or the model’s tendency to exploit spurious
correlations in the data.
In the context of EEG-based classification for epilepsy
diagnosis, models such as DNN and CNN, which show a more
balanced trade-off between training and validation accuracy,
may offer greater clinical reliability. Their performance
aligns with the expectation that validation accuracy should
trail slightly behind training accuracy, reflecting a model’s
ability to generalize while avoiding overfitting. Conversely,
Figure 1. The graphs compare the performance of three models: DNN,
CNN, and LGBM classifier. Both the DNN and CNN models exhibit a the LGBM classifier’s results highlight the importance of
steady increase in training accuracy, eventually plateauing around 98%, rigorous evaluation and regularization in high-stakes clinical
while their validation accuracies stabilize at approximately 80% and 82%, applications, as overly optimistic training performance does
respectively. This is a typical and expected behavior, where the training not necessarily translate to robust diagnostic capabilities.
accuracy is slightly higher than the validation accuracy, reflecting proper These findings underscore the need for careful model
learning without severe overfitting. In contrast, the LGBM classifier
shows a perfect training score of 1.0 across all training instances, which selection and tuning in machine learning-based approaches
is not ideal, as it suggests overfitting on the training data. The cross- to EEG signal classification.
validation score for the LGBM classifier stabilizes at around 85% but lacks
the expected steady increase and plateau shape, which would indicate 3.2. Confusion matrices
better generalization.
Abbreviations: CNN: Convolutional neural network; DNN: Dense Neural A confusion matrix is a widely used ML-based tool for
Network; LightGBM: Light gradient boosting machine. evaluating the performance of classification models. It

Volume 4 Issue 2 (2025) 116 doi: 10.36922/an.7941