CMTS-GNN: a cross-modal temporal-spectral graph neural network with cognitive network explainability.

Area of Science:

• * Neurology and Artificial Intelligence
• * Epilepsy Research
• * Medical Signal Processing

Background:

• * Infantile spasms (IS) are a severe epileptic encephalopathy in infancy, with significant neurodevelopmental consequences if detection is delayed or inaccurate.
• * Conventional electroencephalogram (EEG) analysis for IS is challenging due to data complexity, volume, and inter-rater variability.
• * Accurate and timely IS diagnosis is crucial for mitigating adverse perceptual, cognitive, and affective developmental outcomes.

Purpose of the Study:

• * To develop and validate CMTS-GNN, a Cross-Modal Temporal-Spectral Graph Neural Network, for automated and interpretable detection of infantile spasms from EEG.
• * To improve the accuracy and efficiency of IS detection compared to existing methods.
• * To enhance the interpretability of automated seizure detection models by incorporating brain-region connectivity.

Main Methods:

• * Development of CMTS-GNN, integrating temporal and spectral EEG features using bidirectional cross-modal attention and gated fusion.
• * Incorporation of brain-region connectivity modeling to capture functional interactions.
• * Evaluation on in-house and public CHB-MIT epilepsy datasets using cross-validation and subject-independent schemes.

Main Results:

• * CMTS-GNN achieved high performance on the in-house dataset (e.g., 99.02% accuracy, 98.20% F1-score) and the CHB-MIT dataset (e.g., 98.54% accuracy, 98.47% F1-score).
• * The model demonstrated robustness and generalizability in subject-independent evaluations.
• * CMTS-GNN provided clinically meaningful explanations, enhancing model interpretability.

Conclusions:

• * CMTS-GNN presents an accurate, generalizable, and interpretable framework for automated infantile spasms detection from EEG.
• * The model has the potential to facilitate earlier clinical intervention for IS.
• * Improved detection can help reduce long-term neurodevelopmental morbidity in affected infants.