Explainable Biomedical Graph Transformer for Automated Epilepsy Phenotype Prediction from EHR Data

Research Gap

Existing approaches to epilepsy prediction have been centered around signal processing of EEG based windows or events using traditional machine learning methods and deep neural architectures (SVM, Random Forest, CNN, recurrent, and transformer-based approaches). While these studies increased the accuracy of seizure detection, most of these models relied on hand-engineered signal features and fixed biomedical inputs, and could not address clinical descriptions (primarily physician notes and discharge summaries) present in EHRs. Recent approaches using NLP and ClinicalBERT-based language understanding showed promise in modeling context-specific semantics from physician notes; however, these methods still lack long-range biomedical dependency learning, sufficient explainability, and holistic clinical reasoning of biomedical relationships. Additionally, there is minimal work modeling the relationships between biomedical entities such as seizure activity, medications, EEG results, neuro-checks, and disease progression in a cohesive computational structure.

Figure 1: Research Gap Analysis   Click here to view Figure

Most current models do not employ graph-based explainable reasoning needed for scalable and clinically plausible binary epilepsy phenotype prediction. Figure 1 shows the gap analysis from the literature study.

Materials and Methods

This section focused on end-to-end complete computational framework proposed for the explainable automated prediction of epilepsy phenotypes using Explainable Biomedical Graph Transformer (BioClinX-GraphFormer) illustrated in Figure 2. This framework includes acquiring multimodal health care datasets from MIMIC-IV EHR and CHB-MIT EEG databases to construct an enriched clinical setting to predict epileptic seizures. Preprocessing of unstructured clinical notes includes NLP preprocessing steps such as tokenization, lemmatization, biomedical named entity recognition (NER), and clinical semantics extraction from notes. The clinical entities related to epilepsy like seizure symptoms, seizure medications, EEG descriptions, and neurological manifestations are extracted from clinical notes and encoded to contextual embeddings using Clinical-BERT then structured into biomedical knowledge graph representations. A Graph Transformer Network with multi-head graph attention layers is then built upon this knowledge graph to model long-range semantic and structural dependencies between biomedical concepts. We describe our model training and optimization procedures, evaluation and validation techniques, as well as the explainable attention-based classification process that can be used for reliable, interpretable, and scalable epilepsy prediction in clinical settings.

Figure 2: Flow Diagram of the Proposed Work   Click here to view Figure

Dataset Used

• MIMIC-IV Version 3.1 Dataset⁴²
• CHB-MIT Scalp Dataset⁴³

These two datasets are integrated based on temporal seizure-event mapping and patient-level clinical correlation. EEG seizure annotations from CHB-MIT are aligned with epilepsy-related clinical narratives extracted from MIMIC-IV to create a multimodal biomedical learning environment. The integrated dataset is divided into 70: 15:15 train/validation/test by patient-wise stratified splitting to maintain class balance and avoid data leakage between splits. The primary features for binary epilepsy phenotype prediction task are seizure length, EEG frequency changes, neurologic manifestations, medication usage, clinical events on timeline, physician interpretation, history of diagnosis and biomedical entities relevant to the clinical context from physician notes using Natural Language Processing and ClinicalBERT embeddings.

Clinical Dataset Description and Integration

Data pre-processing converts these heterogeneous clinical notes into biomedical representation that can be used to predict epilepsy phenotype. In our framework, clinical notes provided with Electronic Health Record (EHR) from MIMIC-IV database and seizure annotated electroencephalogram (EEG) data from CHB-MIT database were utilized. Since clinical notes consist of various abbreviations, spelling mistakes, repeated words, time stamps and non-uniform representation of symptoms, rule-based biomedical Natural Language Processing (NLP) pipeline is performed prior to graph construction and transformer learning. Initially each clinical note is normalized by removing non-clinical characters, repeated white spaces, extra punctuations and meaningless words. Also, the conversion of sentences into lower case format, except for major EDA-relevant abbreviations such as EEG, AED, ICU, GTCS, MRI. Let  represent the clinical document. After normalization, the cleaned document is represented as,

On the second stage, tokenization is then performed to split each cleaned clinical narrative into meaningful lexical units. This process separates physician notes, discharge summaries, seizure descriptions, medication terms, and EEG observations into token sequences. The tokenized document is expressed mathematically as,

where, T_i denotes the token sequence and represents the clinical token. Following tokenization, lemmatization is applied to reduce words to their root clinical form. Words like “seizures”, “seizing”, and “seizure-like” map toward representations relating to seizures. This alleviates vocabulary sparsity and increases semantic consistency by:

where, L_i is the lemmatized token sequence. Subsequently, NER extracts epilepsy-related biomedical entities mentioned in the processed text. The NER module captures seizure symptoms, antiepileptic medications, EEG results, neurological symptoms, diagnosis codes, procedures, and time expressions. This can be represented as,

where, E_i is the set of extracted clinical entities. Finally, clinical entity normalization and linking maps extracted entities to standardized biomedical concepts represented in clinical vocabularies like UMLS, SNOMED-CT, ICD codes, RxNorm, etc. For instance, “fits”, “convulsion”, and “seizure episode” are linked to the same seizure concept. This enhances interoperability and decreases ambiguity between heterogeneous records and the equation is,

where, K denotes the biomedical knowledge source and C_irepresents normalized clinical concepts. It contains preprocessed cleaned tokens, clinical terms after lemmatization, detected biomedical entities, normalized Concept Unique Identifiers and time markers of seizures. These features can be utilized to obtain ClinicalBERT embeddings, construct a biomedical knowledge graph and more explainable Graph Transformer-based binary epilepsy phenotype prediction is demonstrated.

Contextual Clinical Embedding and Biomedical Knowledge Graph Construction

The proposed model utilizes Contextual semantic learning and biomedical relationship modeling to forecast epilepsy phenotypes given multimodal Electronic Health Records. After applying clinical entity normalization and data preprocessing steps described above, we represent our normalized clinical notes as semantic embeddings using Clinical-BERT. In contrast to word embedding techniques, Clinical-BERT contextualizes biomedical semantics found in doctors notes, seizure descriptions, administered medication, and neurological exam charts by modeling bidirectional semantic relationships in large-scale medical corpora. Let the preprocessed clinical token sequence be represented as,

where, H_i ∈ R^drepresents the contextual embedding vector of token x_i, and d denotes the embedding dimension. Clinical notes embeddings are aggregated next to produce semantic vector representing the whole clinical document. Semantic document vector stores patterns about seizure evolution, drug use interactions, neurological disorders and physicians’ impressions. Semantic Document vector is computed as,

where, S_d represents the semantic embedding of the clinical document. After obtaining semantic embedding, biomedical entity relationship extraction is conducted to extract semantic relations between seizure-related entities that are meaningful to humans. Seizure symptoms, EEG descriptors, antiepileptic drugs, neurological diseases, diagnostic procedures, and temporal clinical events are extracted as entities. Relation extraction captures relations like “causes”, “associated with”, “treated by”, “indicates”, between biomedical entities based on their semantics. The relation extraction formula is given below,

where, R(E_i, E_j) denotes the semantic relation between biomedical entities E_i and E_j. The extracted entities and relationships are fed into a seizure-centric biomedical knowledge graph. In the knowledge graph, clinical concepts are nodes and biomedical semantic relations are edges. Let’s formally define the knowledge graph as,

where, V represents the set of biomedical nodes and E represents the set of semantic edges connecting clinical entities. The T learning process of node representation produces feature vector for every biomedical entity through contextual embeddings and neighborhood aggregation as follows. Let N(v) represent the neighbor nodes of node v. The node embedding of graph is calculated as,

where, Z_v denotes the learned graph embedding of node v, W₁ and W₂ are trainable weight matrices, and σ represents nonlinear activation. Graph attention learning is integrated within Graph Transformer to capture long-range biomedical semantic dependencies. Attention coefficients within the graph propagation step models which neighboring biomedical entities should be clinically valued more. Formally, the graph attention score can be defined as,

where, α_1j represents the normalized attention coefficient between nodes i and j, and ∥denotes vector concatenation. Due to the constructed biomedical knowledge graph, the proposed framework performs explainable relational reasoning between seizure symptoms, EEG findings, medications, and neurological observations, further enhancing contextualized understanding, semantic dependency modeling and clinical interpretability over traditional machine learning models, sequential BiLSTM and transformer-only epilepsy prediction methods.

Algorithm: Biomedical Graph Construction for Binary Epilepsy Phenotype Prediction

Input: Preprocessed clinical narratives D, EEG annotations E

Output: Biomedical Knowledge Graph Gwith contextual graph embeddings

BioClinX-Graph Former Binary Epilepsy Phenotype Prediction

The Explainable Biomedical Graph Transformer learns to predict epilepsy phenotype with high fidelity. In contrast to standard sequential deep learning approaches, the proposed architecture jointly learns biomedical entity semantic dependency relationships as well as long-distance contextual patterns present in EHR/EEG linked clinical notes. Initially, the biomedical knowledge graph is generated from contextual clinical embeddings is represented as G = (V, E, X), where, V denotes biomedical entity nodes, E represents semantic relationships among entities, and X denotes contextual feature embeddings associated with graph nodes. Biomedical entity nodes are initialized with ClinicalBERT contextual embeddings. The node feature representation is expressed as,

where, h_i⁽⁰⁾represents the initial embedding vector of biomedical concept c_i. After constructing the heterogeneous graph, Graph Transformer Network applies graph attentional methods to propagate relational features. While Graph Convolutional Networks use neighborhood aggregation as computing dependencies, Graph Transformer captures the adaptive semantic interactions between distantly related biomedical entities with self-attention propagation. The attention query, key, and value vectors are computed as,

where, W_Q, W_K, and W_v represent trainable projection matrices. The semantic attention score between biomedical nodes i and j is calculated using scaled dot-product attention which is mathematically expressed as,

where, d_K denotes the dimensionality scaling factor used to stabilize gradient propagation. The normalized graph attention coefficient is computed through Softmax normalization which can be computed as,

where, N(i) denotes neighboring biomedical entities connected to node i. The multi-head graph attention is employed to enhance semantic representation learning. Attention heads learn different biomedical semantics independently, such as seizure evolution, drug-binding, neurological abnormalities, EEG-related abnormalities. The multi-head attention fusion can be formulated as,

where, M represents the number of attention heads and ∥ denotes concatenation. Long-range contextual dependency learning describes another key benefit of our proposed architecture. Seizure indicators can be buried deep within verbose physician notes and discharge summaries. Information tends to degrade when modeling long-range dependencies with traditional recurrent architectures. Global contextual information propagation across graph-connected biomedical entities solves this problem with the Graph Transformer. The contextual dependency propagation is represented as,

where, LayerNorm stabilizes semantic feature propagation across transformer layers. Edge-aware biomedical interaction learning is proposed to enhance relational semantic reasoning. There are semantic edges between seizure-related entities and each edge participates in graph reasoning through edge feature transformation which us represented as,

where, r_ij denotes the relational semantic representation between biomedical entities. The explainable attention-based prediction learns clinically relevant concepts associated with binary epilepsy phenotype prediction. Below are some examples of how the attention weights can be interpreted to reveal important biomedical concepts such as seizure counts, EEG findings, medication resistances and neurological observations etc. Explainability Score is computed as,

where, β_i indicates the clinical importance score of biomedical entity i. Finally, the global graph representation is forwarded into the phenotype classification layer. Here, the graph-level embedding is obtained by applying mean pooling.

where, h_i^(L) represents the final transformer embedding of node i. The Softmax classification is applied to yield the final epilepsy phenotype prediction probability which is represented as,

where, W_c and b_c represent trainable classification parameters.

Figure 3: BioClinX-GraphFormer Clinical NLP Model for Epilepsy Prediction   Click here to view Figure

The proposed model is illustrated in Figure 3 which improves conventional epilepsy prediction models in four aspects: 1) it learns contextual biomedical semantics with ClinicalBERT embeddings; 2) learns relational interactions between seizure-associated entities using graph reasoning; 3) enhance long-range dependency and contextual interpretability with multi-head graph attention; 4) allow clinicians to recognize which biomedical entities are primarily contributing towards prediction through explainable attention.

Model Training and Hyperparameter Optimization

The Explainable Biomedical Graph Transformer (BioClinX-GraphFormer) architecture utilizes state-of-the-art model training and optimization techniques for effective epilepsy prediction. During biomedical learning at scale, it allows for convergence stability, generalizability enhancement, and mitigation of overfitting. Techniques such as Adam optimization, cross entropy loss, dropout regularization, layer normalization, stratified data splitting, and k-fold cross validation comprise the model training pipeline which enables trustworthy clinical predictions. The combined MIMIC – IV and CHB – MIT datasets are split into train, validation, and test subsets before preprocessing using stratified sampling methods. Stratified sampling ensures that class proportions for the epilepsy phenotypes are maintained and class imbalance is avoided during model training. Let the entire dataset be represented as,

where, x_i denotes the biomedical graph representation and y_i represents the corresponding epilepsy phenotype label. The dataset is partitioned into,

using a 70:15:15 ratio for training, validation, and testing respectively. Here, k-fold cross validation is used to increase model robustness and decrease variance. The training data is split into k mutually exclusive folds. Training-validation cycles are run iteratively. Validation accuracy is averaged over folds as follows,

where, Accuracy_i represents the validation accuracy obtained in the i^th fold. Adam optimizer was chosen to train Graph Transformer parameters because of its adaptive learning rate estimation and convergence properties for high-dimensional biomedical graph learning applications. Adam is essentially momentum based gradient descent with adaptive learning rates. The first-order moment estimation is computed as,

where, g_t denotes the gradient at iteration t, and β₁ represents exponential decay. Similarly, the second-order moment estimation is calculated as,

where, β₂ controls variance smoothing. The final parameter update rule is expressed as,

where, η denotes the learning rate and ∈ prevents numerical instability. Since binary epilepsy phenotype prediction can be formulated as a classification problem, the model prediction error is minimized with categorical cross-entropy loss. The loss function is given by,

where, C denotes the number of phenotype classes, Y_i is the actual label, ȳ_i and represents predicted probability. Dropout regularization is added between transformer layers and fully connected layers to avoid overfitting and increase generalization ability. Neurons are randomly dropped out at a rate of  while training the network. The equation for dropout is as follows,

where, r_i ∼ Bernoulli(1 – p). Layer normalization is also adopted to normalize feature propagation and speed up training convergence. Note that layer normalization is performed independently across feature dimension and thus is also appropriate for graph based contextual learning, in contrast to batch normalization. The layer normalized feature representation is given by,

where, μ and σ² represent mean and variance respectively. Hyperparameter optimization helps to increase epilepsy prediction performance. Our framework optimizes,

• Learning rate
• Number of graph attention heads
• Transformer depth
• Hidden embedding dimension
• Dropout probability
• Batch size Weight decay coefficient

using iterative validation-driven search. The set of hyperparameters that yield the best validation accuracy and convergence to minimum loss is chosen.

Criteria for early stopping is included to check for convergence of training. Training stops when the validation loss hasn’t decreased in a certain number of epochs. This is given by,

where, δ denotes convergence tolerance. Preventing Overfitting is further accomplished with,

• Dropout regularization
• Early stopping
• Layer Normalization
• Attention Regularization
• Cross-Validation Averaging

Together, this provides our Explainable Biomedical Graph Transformer model the ability to accurately train biomedical semantics and relation embeddings with high clinical interpretability and trustworthiness for automated epilepsy phenotype predictions.

Automated Binary Epilepsy Phenotype Classification and Performance Evaluation

The proposed BioClinX-GraphFormer is used for automated binary epilepsy classification. The model learns contextual biomedical semantics, models biomedical relationships between clinical entities, and provides explainable decision support using graph attention scores. Clinical attributes such as seizure manifestations, EEG findings, current medications, neuro exam, disease progression, and clinical notes are represented as graph nodes and propagated through explainable Graph Transformer layers for context-aware node feature aggregation. Graph embeddings produced by BioClinX-GraphFormer encapsulate semantic as well as relational clinical information for complete contextual representation of patient evidence. Lastly, graph embeddings from the transformer layer served as input to a Softmax classifier for binary classification task (Epilepsy vs. Non-Epilepsy).

Initially, the graph-level embedding generated from the Graph Transformer architecture is represented as,

where, G_final denotes the final contextual biomedical feature representation extracted from seizure-centric biomedical graphs. The dense classification layer projects the learned embeddings onto epilepsy specific logits corresponding to classes of,

• Focal Onset Epilepsy
• Generalized Epilepsy
• Combined Generalized and Focal Epilepsy
• Unknown Epilepsy Type

The logit computation is represented as,

where W_i and b_i represent trainable weight and bias parameters associated with phenotype class i. The Softmax activation function converts logits into normalized probability distributions which can be computed as,

where, P(y_i) denotes the probability of epilepsy phenotype i, and c represents the total number of phenotype categories. The phenotype class with the highest probability is selected as the final prediction:

Discussion

All traditional machine learning and deep learning baselines were outperformed by our proposed BioClinX-GraphFormer model on every metric shown in Table 4. Models like SVM, Random Forest, and XGBoost were limited by their ability to understand context due to using hand-engineered biomedical features and handcrafted representations of EEG features. NLP-BiLSTM provided enhanced learning of sequential context from the patient’s EHR text notes. Transformer-Enhanced ClinicalBERT provided improved learning of long-range semantic dependencies using self-attention layers. Our proposed model builds upon these methods by integrating,

• ClinicalBERT contextual embeddings
• Biomedical knowledge graph reasoning.
• Multi-head graph attention.
• Explainable semantic inference.

This allows for accurate modeling of biomedical relationships between seizure symptoms, EEG findings, medication, and neuro checks. This led to fewer false predictions and higher clinically interpretable epilepsy predictions. For the binary Epilepsy vs. Non-Epilepsy and Seizure vs. Non-Seizure tasks, BioClinX-GraphFormer jointly conditions on contextual EHR data, seizure relevant EEG attributes, mentions of medications, and neurological notes.

Training stability is additionally measured to monitor how smoothly the optimization process behaves during training by calculating the variance in the training loss over epochs. Lower TS indicates smoother convergence, more stable parameter updates, and better optimization reliability. BioClinX-GraphFormer achieves better training stability with graph transformer learning, layer normalization, multi-head graph attention and Adam optimizer to allow binary epilepsy classification from multimodal EHR and EEG data to converge steadily.

The performance of BioClinX-GraphFormer reported is limited to MIMIC-IV and CHB-MIT datasets and the experimental protocol outlined in this paper. Therefore, these results should not be interpreted as evidence of general clinical performance, and further validation on independent, multi-centre datasets is required to establish the robustness and generalisability of the proposed framework.

Advantages and Disadvantages of the Proposed Methodology

Potential Contributions compared with existing related works are, the proposed model is capable of providing a number of benefits to address the challenges associated with automated binary epilepsy prediction. ClinicalBERT contextual embeddings coupled with biomedical knowledge graph learning permit semantic contextualization of high-dimensional EHR as well as seizure-phenotype relevant clinical texts. Multi-head graph attention allows modeling long-range dependencies between biomedical entities (e.g seizure symptoms, EEG aberrations, pharmaceutical terms, neurology observations etc.). Explainable attention mechanism allows our framework to generate clinically interpretable predictions which can be transparently understood by physicians. Graph-based neural representation learning can capture nonlinear relationships between EHR features and permits state-of-the-art prediction accuracy, fewer false positives/negatives, and scalability to large-scale health care settings with multimodal biomedical data.

Potential limitations of proposed methodology are computationally expensive and requires additional GPU memory bandwidth when applied to large-scale graph learning problems. Generation of Knowledge Graphs and Biomedical entity embeddings require additional time for pre-processing step. This framework also relies on availability of high-quality clinical annotations and entity extraction. Implementation of model in real-time scenarios with limited resources might require model compression and knowledge distillation to reduce the compute costs.

Future Work

Future work will aim to address multimodal biomedical fusion for personalized epilepsy prediction by learning from real-time EEG streams, neuroimages, genomic biomarkers and wearable healthcare sensors. In addition, the model will still investigate federated learning and lightweight Graph Transformer pruning/inference techniques for privacy preserving, resource efficient clinical deployment. Future models will also leverage temporal disease progression modeling, adaptive graph reasoning and explainable multimodal clinical decision-support for early seizure alerting, intervention and treatment recommendations as well as real-time neurological healthcare intelligence.

Conclusion

In this study, a hybrid Explainable Biomedical Graph Transformer (BioClinX-GraphFormer) was proposed for binary epilepsy/seizure classification using EHRs and seizure-related EEG information. The framework integrates ClinicalBERT contextual embeddings, biomedical knowledge graph construction, Graph Transformer learning, multi-head graph attention, and explainable semantic reasoning to jointly model contextual and relational clinical information. Modeling long-range semantic dependencies and biomedical relationships between seizure manifestations, seizure medications, EEG abnormalities, and neurologic observations allows for better feature representation than traditional machine learning and sequential DL approaches. Experimental results on benchmark MIMIC-IV and CHB-MIT datasets resulted in prediction accuracy of 98.1%, F1- score of 98.0%, and AUROC of 0.985 with fewer false positives and false negatives seizures under our experimental setting. Utilizing the graph attention mechanism improves interpretability of the automated prediction by visualizing contributing biomedical entities and relationships. The proposed model offers improvements to the accuracy and interpretability of intelligent epileptic prediction and can serve as foundational work for future Explainable Biomedical Artificial Intelligence applications and real-time clinical decision-making tasks.

Acknowledgement

The authors would like to thank the research institution Dr. SNS Rajalakshmi College of Arts & Science, Coimbatore (Tamil Nadu) for providing adequate facility and infrastructure to complete this research work in a timely manner. The authors also acknowledged that the usage of modern infrastructure like computational biology tools and libraries during the implementation process of this research work. TensorFlow, Scikit-learn, Google Colab and MATLAB were used to evaluate experiments. Mermaid Live Editor is used for flow diagrams and workflows to represent the research model. Grammarly was used to improve grammar and language fluency. The authors themselves conducted all experiments, methodology, model designs, and analyses.

Funding Sources

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Conflict of Interest

The authors do not have any conflict of interest.

Data Availability Statement

This statement does not apply to this article.

Ethics Statement

This research did not involve human participants, animal subjects, or any material that requires ethical approval.

Informed Consent Statement

This study did not involve human participants, and therefore, informed consent was not required.

Clinical Trial Registration

This research does not involve any clinical trials.

Permission to reproduce material from other sources

Not Applicable.

Author Contribution

• Leeshma Koroth: Conceptualization, Methodology, Writing – Original Draft, Data Collection, Analysis, Writing – Review & Editing
• Praveena Marannan: Visualization, Supervision, Project Administration

References

1. Ali E., Angelova M., Karmakar C. Epileptic Seizure Detection Using CHB-MIT Dataset: The Overlooked Perspectives. Royal Society Open Science. 2024;11(5):230601. https://doi.org/10.1098/rsos.230601
   CrossRef
2. Atitallah S. B., Driss M., Boulila W., Koubaa A. Graph-Based EEG Analysis for Seizure Prediction Enhanced with Kolmogorov–Arnold Networks and Self-Supervised Learning. Engineering Science and Technology an International Journal. 2025;73:102245. https://doi.org/10.1016/j.jestch.2025.102245
   CrossRef
3. Alharthi M. K., Moria K. M., Alghazzawi D. M., Tayeb H. O. Epileptic Disorder Detection of Seizures Using EEG Signals. 2022;22(17):6592. https://doi.org/10.3390/s22176592
   CrossRef
4. Tan S., Tang Z., He Q., et al. Automatic Detection and Prediction of Epileptic EEG Signals Based on Nonlinear Dynamics and Deep Learning: A Review. Frontiers in Neuroscience. 2025;19:1630664. https://doi.org/10.3389/fnins.2025.1630664
   CrossRef
5. Leeshma K., Praveena M. Natural Language Processing and Machine Learning Based Automated Diagnosis of Epilepsy in Electronic Health Record. IEEE International Conference on Computing, Sciences and Communications (ICCSC). 2024:1-6. https://doi.org/10.1109/ICCSC62048.2024.10830353
   CrossRef
6. Atlam H. F., Aderibigbe G. E., Nadeem M. S. Effective Epileptic Seizure Detection with Hybrid Feature Selection and SMOTE-Based Data Balancing Using SVM Classifier. Applied Sciences. 2025;15(9):4690. https://doi.org/10.3390/app15094690
   CrossRef
7. Zhao X., Yoshida N., Ueda T., Sugano H., Tanaka T. Epileptic Seizure Detection by Using Interpretable Machine Learning Models. Journal of Neural Engineering. 2023;20(1):015002. https://doi.org/10.1088/1741-2552/acb089
   CrossRef
8. Betgeri S. S., Shukla M., Kumar D., Khan S. B., Khan M. A., Alkhaldi N. A. Enhancing Seizure Detection with Hybrid XGBoost and Recurrent Neural Networks. Neuroscience Informatics. 2025;100206. https://doi.org/10.1016/j.neuri.2025.100206
   CrossRef
9. Leeshma K., Praveena M. Natural Language Processing Based BiLSTM Model for Automated Epilepsy Diagnosis from Electronic Health Records. Indian Journal of Science and Technology. 2026;19(19):1225-1240. https://doi.org/10.17485/IJST/v19i19.1317
   CrossRef
10. Xie K., Gallagher R. S., Conrad E. C., et al. Extracting seizure frequency from epilepsy clinic notes: a machine reading approach to natural language processing, Journal of the American Medical Informatics Association, 2022; 209(5):873–881. https://doi.org/10.1093/jamia/ocac018
    CrossRef
11. Eldho K. J., Nithyanandh S. Lung Cancer Detection and Severity Analysis with a 3D Deep Learning CNN Model Using CT-DICOM Clinical Dataset. Indian Journal of Science and Technology. 2024;17(10):899-910. https://doi.org/10.17485/IJST/v17i10.3085
    CrossRef
12. Maeda, K., Tsuboi, H., Hosoda, N., et al. Phenotypic Classification of Scalp High-Frequency Oscillations in Absence Epilepsy Based on Multiple Characteristics Using K-Means Clustering. 2026;13(1):65. https://doi.org/10.3390/bioengineering13010065
    CrossRef
13. Hirsch E., French J., Scheffer I. E., Bogacz A., Alsaadi T., Sperling M. R. ILAE Definition of the Idiopathic Generalized Epilepsy Syndromes: Position Statement by the ILAE Task Force on Nosology and Definitions. 2022;63:1475-1499. https://doi.org/10.1111/epi.17236
    CrossRef
14. Krikid F., Karfoul A., Chaibi S., et al. Multi-Classification of High-Frequency Oscillations in Intracranial EEG Signals Based on CNN and Data Augmentation. Signal Image and Video Processing. 2024;18:1099-1109. https://doi.org/10.1007/s11760-023-02808-4
    CrossRef
15. Zhang Y., Lu Q., Monsoor T., et al. Refining Epileptogenic High-Frequency Oscillations Using Deep Learning: A Reverse Engineering Approach. Brain Communications. 2022;4(1):fcab267. https://doi.org/10.1093/braincomms/fcab267
    CrossRef
16. Namasse Z., Tabaa M., Hidila Z., Mouchawrab S. Explainable Artificial Intelligence for Predicting Attention Deficit Hyperactivity Disorder in Children and Adults. 2025;13(2):155. https://doi.org/10.3390/healthcare13020155
    CrossRef
17. Kunekar P., Kumawat C., Lande V., Lokhande S., Mandhana R., Kshirsagar M. Comparison of Different Machine Learning Algorithms to Classify Epilepsy Seizure from EEG Signals. Engineering Proceedings. 2023;59(1):166. https://doi.org/10.3390/engproc2023059166
    CrossRef
18. Indhumathi G., Anil P. S., Posiyya A., Nithyanandh S., Suresh H. R., Navaneethan S. Deep Learning-Based Tongue Biometrics for Secure Authentication in IoT-Driven Healthcare Systems. International Conference on Smart & Sustainable Technology (INCSST). 2025:1-6. https://doi.org/10.1109/INCSST64791.2025.11210319
    CrossRef
19. Karasmanoglou A., Antonakakis M., Zervakis M. ECG-Based Semi-Supervised Anomaly Detection for Early Detection and Monitoring of Epileptic Seizures. International Journal of Environmental Research and Public Health. 2023;20(6):5000. https://doi.org/10.3390/ijerph20065000
    CrossRef
20. Farooq M. S., Zulfiqar A., Riaz S. Epileptic Seizure Detection Using Machine Learning: Taxonomy, Opportunities, and Challenges. 2023;13(6):1058. https://doi.org/10.3390/diagnostics13061058
    CrossRef
21. Huang X., Sun X., Zhang L., et al. A Novel Epilepsy Detection Method Based on Feature Extraction by Deep Autoencoder on EEG Signal. International Journal of Environmental Research and Public Health. 2022;19(22):15110. https://doi.org/10.3390/ijerph192215110
    CrossRef
22. Aslam M. H., Usman S. M., Khalid S., et al. Classification of EEG Signals for Prediction of Epileptic Seizures. Applied Sciences. 2022;12(14):7251. https://doi.org/10.3390/app12147251
    CrossRef
23. Jaishankar B., A. M. A., D. V., Raja L. A. Novel Epilepsy Seizure Prediction Model Using Deep Learning and Classification. Healthcare Analytics. 2023;4:100222. https://doi.org/10.1016/j.health.2023.100222
    CrossRef
24. Khan M. A. R., Heba K. M., Abbass A. H. An Epilepsy Prediction and Management System Based on Federated Learning Combined with Hybrid Harmony Search and Mutual Information (HSA-MI)-Based Feature Selection Approach. Scientific Reports. 2026;16:893. https://doi.org/10.1038/s41598-025-30919-0
    CrossRef
25. Shekhar J., Ramya M., Chandrasekar A., Nithyanandh S., D. P., Madhan A. MiniGCN: Mini-Batch Graph Convolutional Networks for Hyperspectral Image Classification. Global Conference on Information Technology and Communication Networks (GITCON). 2025:1-6. https://doi.org/10.1109/GITCON65266.2025.11378144
    CrossRef
26. Song Z., Arrowsmith K., Henare D., Pedersen M. Real-Time EEG-Based Epileptic Seizure Prediction Using Artificial Intelligence: A Systematic Review. 2025. https://doi.org/10.1101/2025.10.09.25337692
    CrossRef
27. Zhu R., Pan W. X., Liu J. X. Epileptic Seizure Prediction via Multidimensional Transformer and Recurrent Neural Network Fusion. Journal of Translational Medicine. 2024;22:895. https://doi.org/10.1186/s12967-024-05678-7
    CrossRef
28. Jayanti R., Jain T. Epileptic Seizure Detection and Prediction from EEG Data: A Machine Learning Approach with Clinical Validation. 2025. 10.48550/arxiv.2510.24986
29. Qiao W., Bi X., Han L., Zhang Y. Epilepsy Prediction and Detection Using Attention-CssCDBN with Dual-Task Learning. 2025;25(1):51. https://doi.org/10.3390/s25010051
    CrossRef
30. Mgohele R. N., Massawe B. H. J., Shitindi M. J., Sanga H. G., Omar M. M. Prediction of Soil Texture Using Remote Sensing Data: A Systematic Review. Frontiers in Remote Sensing. 2024;5:1461537. https://doi.org/10.3389/frsen.2024.1461537
    CrossRef
31. S., Vaishnavi S., T. S. A., Nithyanandh S., M. S., Nithya P. Energy-Efficient WaOA-CNN Hybrid Model for Medical Image Classification in Edge Healthcare. International Conference on Communication, Computing and Emerging Technologies (IC3ET). 2026:582-587. https://doi.org/10.1109/IC3ET64989.2026.11467237
    CrossRef
32. Huang Y., Li Q., Yang Q., et al. Early Prediction of Refractory Epilepsy in Children Under Artificial Intelligence Neural Network. Frontiers in Neurorobotics. 2021;15:690220. https://doi.org/10.3389/fnbot.2021.690220
    CrossRef
33. Linden T., De Jong J., Lu C., Kiri V., Haeffs K., Fröhlich H. An Explainable Multimodal Neural Network Architecture for Predicting Epilepsy Comorbidities Based on Administrative Claims Data. Frontiers in Artificial Intelligence. 2021;4:610197. https://doi.org/10.3389/frai.2021.610197
    CrossRef
34. Ibrahim A. K., Zhuang H., Tognoli E., Muhamed Ali A., Erdol N. Epileptic Seizure Prediction Based on Multiresolution Convolutional Neural Networks. Frontiers in Signal Processing. 2023;3:1175305. https://doi.org/10.3389/frsip.2023.1175305
    CrossRef
35. Yang Y., Qin X., Wen H., Li F., Lin X. Patient-Specific Approach Using Data Fusion and Adversarial Training for Epileptic Seizure Prediction. Frontiers in Computational Neuroscience. 2023;17:1172987. https://doi.org/10.3389/fncom.2023.1172987
    CrossRef
36. Huo Q., Luo X., Xu Z. C., Yang X. Y. Machine Learning Applied to Epilepsy: Bibliometric and Visual Analysis from 2004 to 2023. Frontiers in Neurology. 2024;15:1374443. https://doi.org/10.3389/fneur.2024.1374443
    CrossRef
37. Bangash A. H., Bercu M. M., Byrne R. W., Pavuluri S., Salehi A. Application of Machine Learning Approaches to Predict Seizure-Onset Zones in Patients with Drug-Resistant Epilepsy: A Systematic Review. Frontiers in Neurology. 2025;16:1687144. https://doi.org/10.3389/fneur.2025.1687144
    CrossRef
38. Bosl W. J., Leviton A., Loddenkemper T. Prediction of Seizure Recurrence: A Note of Caution. Frontiers in Neurology. 2021;12:675728. https://doi.org/10.3389/fneur.2021.675728
    CrossRef
39. Flegel T., Neumann A., Holst A. L., et al. Machine Learning Algorithms Predict Canine Structural Epilepsy with High Accuracy. Frontiers in Veterinary Science. 2024;11:1406107. https://doi.org/10.3389/fvets.2024.1406107
    CrossRef
40. Selvam N., Joy J. K. Plant Leaf Disease Detection with Multivariable Feature Selection Using Deep Learning AEN and Mask R-CNN in PLANT-DOC Data. Biotechnology Research Asia. 2024;21(4). http://dx.doi.org/10.13005/bbra/3333
    CrossRef
41. Shaby S. M., Vasuki N., Priya R. A., Nithyanandh S., Sornavalli M., Krishnakumari S. AI-Powered Low-Dose CT Reconstruction Using Autoencoders for Radiation Risk Reduction. International Conference on Electronic Systems and Intelligent Computing (ICESIC). 2026:941-946. https://doi.org/10.1109/ICESIC67389.2026.11496448
    CrossRef
42. Johnson A, Bulgarelli L, Pollard T, et al. MIMIC-IV (version 3.1). PhysioNet. 2024. https://doi.org/10.13026/kpb9-mt58
43. Guttag J. CHB-MIT Scalp EEG Database (version 1.0.0). PhysioNet. 2010. https://doi.org/10.13026/C2K01R