Page 119 - AN-4-2

Advanced Neurology ML for EEG signal recognition

can either present with impaired awareness or without it. patient outcomes by potentially reducing side effects and
4
Those with reduced awareness involve a change or loss of enhancing seizure control. In addition, timely screening
consciousness, where individuals may seem to be awake but can help prevent the progression of the condition, reducing
unresponsive, often engaging in repetitive movements. In the risk of complications and improving overall quality of
contrast, focal seizures without impaired awareness can alter life for patients. 14
emotions and sensory perceptions without causing a loss Recent developments in deep learning architectures,
of consciousness, leading to symptoms such as emotional particularly those integrating temporal and spectral analysis
changes, sensory disturbances, or physical jerking. techniques, have shown remarkable accuracy (up to 100%
Generalized seizures affect all areas of the brain from the in binary classifications) for seizure detection. 15-18 This
moment they begin, causing widespread disruption in brain underscores the significant potential of combining various
activity. There are various types of generalized seizures analytical approaches in enhancing detection capabilities.
5
including absence seizures, tonic seizures, clonic seizures, However, while studies reporting exceptional classification
and myoclonic seizures.
metrics highlight methodological promise, many lack
Epilepsy can be diagnosed through various methods, comprehensive validation protocols to ensure clinical
including a range of imaging techniques. These methods help robustness. For instance, claims of near-perfect accuracy
assess brain activity and pinpoint abnormalities, contributing often omit critical evidence of model training dynamics
to a comprehensive diagnosis of the condition. The most (e.g., learning/loss curves) or class separation reliability,
common imaging technique used to diagnose epilepsy is such as the balance between true positive rates (TPR,
the electroencephalogram (EEG). An EEG measures the i.e., correctly identified seizures) and false positive rates
6,7
electrical activity in various parts of the brain by placing (FPR, i.e., erroneous seizure alerts) across classification
small electrodes on the scalp to detect abnormalities in brain thresholds (i.e., receiver operating characteristic [ROC]
wave patterns. Abnormal brain wave patterns detected curves), making it difficult to assess whether reported
8
on an EEG can assist doctors in determining the type of performance reflects generalizable learning or dataset-
seizure and identifying the specific area of the brain affected. specific overfitting. Moreover, in the context of assessing
19
These irregularities are key in diagnosing epilepsy and these learning architectures for tasks like seizure detection,
understanding its origins within the brain. High-density EEG it is important to compare training versus validation
is a variation of the standard EEG where electrodes are placed learning/loss curves, which many of these studies omit.
closer together, allowing for more precise identification of This comparison provides insights into the model’s training
the brain areas affected by seizures. 9 dynamics, allowing the reader to observe whether the curves

Diagnosing epilepsy involves several steps: first, converge, merge, and eventually plateau. Such behaviors
determining if the episode was an epileptic seizure or are indicative of a model that is learning effectively and
another event, such as syncope or cardiac arrhythmias, generalizing well to unseen data. Without demonstrating
which tend to present with similar symptoms. Once a these dynamics, one cannot ascertain whether reported
10
seizure is confirmed, the next step is to classify the seizure high accuracy (such as claims of up to 100% in binary
type, followed by identifying the underlying cause through classifications) holds any meaningful clinical significance.
additional tests such as EEGs, computed tomography (CT) If the learning curves show divergence or erratic behavior,
scans, or blood tests. The final step involves classifying the it may suggest overfitting to the training data rather than
epilepsy syndrome, which helps guide treatment choices, true generalization, thereby undermining the reliability
19
including the selection of antiepileptic drugs. This process of the model in real-world clinical applications. Thus,
can take weeks to months, depending on the complexity transparent evaluation frameworks that incorporate both
of the case and the need for further investigations. The performance metrics and convergence behavior analyses
diagnosis is often complicated by interobserver variability, are essential for translating algorithmic advancements into
as different doctors may interpret clinical descriptions and clinically robust solutions. This underscores the need for
tests differently, leading to delays or errors. 11,12 transparent evaluation frameworks that pair performance
metrics with convergence behavior analyses when
Overall, more efficient screening for epilepsy can lead translating algorithms to clinical settings. 20
to faster and more accurate diagnoses, allowing for earlier
intervention and better-targeted treatment. By identifying 2. Materials and methods
13
the specific type of epilepsy or seizure syndrome quickly,
healthcare providers can tailor treatment plans more 2.1. Data description
precisely, reducing the trial-and-error approach that The dataset utilized for this study was derived from Kaggle,
often comes with antiepileptic drugs. This can improve an open-source online database. It consists of 2,216 rows,
21

Volume 4 Issue 2 (2025) 113 doi: 10.36922/an.7941