Page 27 - AN-4-3

Advanced Neurology Artificial intelligence in epilepsy education

their primary applications, strengths, and limitations. This reporting an accuracy of 92.1%, sensitivity of 90.5%, and
table aims to give a concise overview of each model’s role and specificity of 93.3%, which demonstrates solid predictive
applicability in addressing the unique challenges presented capability, although it falls short of the advanced accuracy
in epilepsy care. Table 1 shows the characteristic features of levels achieved by models like ResBiLSTM. 56
different AI applications in epilepsy management. These studies collectively illustrate the strengths

Several studies have explored machine learning and limitations of various machine learning models
models for detecting epileptic seizures based on EEG data, in epilepsy detection. 49-57 Although CNN and LSTM
achieving varying degrees of accuracy, sensitivity, and models consistently demonstrate high sensitivity and
specificity. 48-56 Hannun et al. developed a CNN for seizure specificity, their performance can vary based on dataset
45
detection, achieving an accuracy of 95.6%, sensitivity of characteristics and model refinements. 49-51 Models, like
93.2%, and specificity of 96.5%. However, when compared ResBiLSTM, show that integrating bidirectional temporal
47
with the lightweight triscale yielding-CNN model by processing can enhance accuracy, especially in binary
Yang et al., which attained an accuracy of 99.9% on the classification. The effectiveness of each model appears
56
48
SWEC-ETHZ dataset, Rajpurkar’s CNN shows slightly linked to specific factors, such as EEG data type, temporal
reduced accuracy while still being a practical approach analysis, and feature extraction techniques, underscoring
45
for seizure detection. Similarly, Mirowski et al. applied the importance of selecting models based on the dataset
49
support vector machines (SVM) for heart rate-based and clinical application requirements. 56,58 This comparative
seizure detection and obtained an accuracy of 91.4%, with analysis highlights the promise of refined neural network
a sensitivity of 89.2%, and a specificity of 92.5%. This was architectures, particularly for achieving near-perfect
49
lower than the results reported by Pedersen et al., who accuracy in seizure detection. Table 2 shows an overview
50
also used SVM with relative energy features from discrete of AI models currently utilized for epilepsy diagnosis and
wavelet transform, achieving a higher accuracy of 98%. 50 prediction.
In a different approach, Balta used a random forest 8. Leading institutions in AI epilepsy
51
classifier for early seizure detection on EEG data, achieving research
an accuracy of 90.8%, sensitivity of 88.5%, and specificity
of 91.9%. In contrast, Huang achieved over 96% Institutions worldwide are making significant strides
51
52
accuracy using a random forest with intracranial EEG data, in applying AI technologies to improve epilepsy
showing that model performance can vary substantially management. 58-60 The National Institute of Neurological
across different datasets. Pascual et al., used a long Disorders and Stroke is at the forefront, conducting
52
53
short-term memory (LSTM) model to capture temporal pioneering research in CNNs to enhance seizure prediction
dependencies within EEG signals, reporting an accuracy capabilities. 58-60 The Epilepsy Foundation is actively
53
of 94.2%, sensitivity of 92.1%, and specificity of 95.2%. funding projects focused on AI-driven seizure detection
In comparison, Zhao et al. implemented a residual to introduce innovative solutions into clinical practice. 58,59
56
bidirectional LSTM (ResBiLSTM) model, achieving up At the University of California, Los Angeles, researchers
to 100% accuracy in binary classification, indicating a are developing RNN-based diagnostic tools that promise
marked improvement over Zhang’s approach, particularly earlier and more accurate epilepsy detection. 58,59 The
in managing temporal dependencies. Lastly, Mbiazi University of Oxford is advancing AI-powered EEG
56
et al. deployed a RNN model for seizure prediction, analysis, which is essential for precise seizure monitoring.
57
60
Table 1. Comparative analysis of AI applications in epilepsy management by type of model and purpose

AI model Primary application Advantages Limitations
CNN 45 EEG signal classification High accuracy in pattern recognition Computationally intensive
RNN 45,46 Seizure prediction and temporal Suitable for time-series data Requires large datasets
analysis
SVM 47 Feature-based EEG classification Effective for small datasets Limited flexibility with complex data
Decision trees 47 Patient outcome prediction Easy interpretability Prone to overfitting in high-dimensional data
Transfer learning 48 Adaptation to specific patient groups Reduces the need for extensive training data Requires pre-trained, similar domain models
Reinforcement Personalized treatment and adherence Dynamic adaptation to patient needs Complex to train and validate in
learning 47,48 healthcare
Abbreviations: CNN: Convolutional neural network; EEG: Electroencephalography; RNN: Recurrent neural networks; SVM: Support vector machines.

Volume 4 Issue 3 (2025) 21 doi: 10.36922/an.4777