Page 126 - AN-4-2

Advanced Neurology ML for EEG signal recognition

While our results demonstrate that CNNs inherently ability to capture spatial and temporal features, can serve
amplify neurophysiologically discriminative features in as a reliable foundation for automated epilepsy screening
intermediate layers (e.g., spectral asymmetries), their tools. Such tools could enable faster and more accurate
decision-making process requires explicit linkage to diagnoses, reducing the burden on neurologists and
clinician-annotated biomarkers. Recent studies employing improving patient outcomes through earlier intervention.
permutation entropy analysis have shown that CNNs The balanced performance across both epileptic and non-
trained on EEG data progressively enhance class separability epileptic cases also minimizes the risk of misdiagnoses,
through sequential convolutional layers, suggesting learned which is crucial for maintaining trust in clinical settings.
29
filters align with neuroanatomical seizure patterns. Future To further enhance clinical utility, additional efforts should
28
implementations will integrate post hoc interpretability focus on improving model interpretability and validating
tools such as gradient-weighted class activation mapping these approaches across diverse patient populations
to visualize spatiotemporal regions driving predictions and clinical environments. For CNNs, incorporating
30
(e.g., lateralized interictal spikes) and Shapley additive advanced architectures, such as attention mechanisms or
explanations to quantify feature importance across recurrent layers, may enhance their ability to capture long-
spectral bands. These methods will explicitly map model range temporal dependencies in EEG data. Expanding
31
outputs to established epileptiform criteria (e.g., polyspike the dataset and incorporating more diverse samples
morphology, ictal rhythms), enabling clinicians to validate could help reduce bias and improve generalization across
predictions against domain knowledge while preserving patient populations. While CNNs demonstrated the most
32
classification accuracy. Such interpretability not only promise for EEG-based epilepsy detection, optimizing
validates CNN reliability but also bridges the gap between each model for its specific strengths could further enhance
“black-box” deep learning and clinical trustworthiness, diagnostic accuracy and reliability.
ensuring model decisions reflect neurophysiologically This study employed a single publicly available EEG
meaningful patterns rather than spurious correlations.
dataset to ensure methodological consistency during
While LightGBM performed well on validation data, its comparative model evaluation. While this approach
lack of convergence highlights potential overfitting. CNNs controlled for confounding variables, we acknowledge
exhibited only a slight gap between training and validation that generalization across diverse clinical scenarios (such
performance, indicating a much greater generalizability. as pediatric vs. adult populations, focal vs. generalized
DNNs struggled to match the performance of CNNs due to seizure types, or EEG recordings from varying hardware)
their inability to leverage spatial and temporal dependencies. requires validation on multicenter, demographically
This limitation underscores the importance of selecting heterogeneous datasets. EEG signal characteristics
32
architectures that align with the data’s underlying structure. inherently vary with age, pharmacological history, and
The high dimensionality and temporal characteristics comorbid neurological conditions, which are factors not
of EEG signals pose unique challenges for ML models. fully captured in our dataset. Future work will expand to
CNNs’ superior performance reflects their strength in multicenter collaborations incorporating wearable EEG
addressing these challenges, but even they are not immune devices, which capture longitudinal data across real-world
to misclassifications, particularly in borderline cases. environments (e.g., sleep, stress).
CNNs’ ability to consistently outperform other models
suggests their potential as a reliable tool for epilepsy 5. Conclusion
detection using EEG signals. Their robustness in handling This study highlights the potential of ML models,
complex patterns ensures higher diagnostic accuracy, particularly CNNs, for classifying epileptic and non-
which is critical for timely intervention. The ROC curves epileptic EEG signals. CNNs outperformed other models in
for both the CNN and DNN models demonstrate strong this study due to their superior ability to capture complex
performance in distinguishing epileptic from non- spatial and temporal patterns, making them well-suited for
epileptic EEG signals. The small differences between analyzing non-stationary EEG data. Critically, their training
validation and testing datasets highlight the robustness dynamics, evidenced by convergent training/validation
of the models, making them promising candidates for learning curves and stable ROC metrics, demonstrate
clinical applications. However, further validation on larger robust generalization, addressing concerns that high
and more diverse datasets is recommended to ensure their accuracy in prior studies might stem from dataset-specific
reliability in real-world scenarios. overfitting rather than clinically transferable learning.
The implications of these findings in clinical While LightGBM exhibited high initial accuracy and
applications are significant. CNNs, with their superior precision, its lack of convergence (i.e., divergent training-

Volume 4 Issue 2 (2025) 120 doi: 10.36922/an.7941