LLM-Powered Cross-Modal Alignment for Explainable Seizure Detection from EEG.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Artificial intelligence (AI) has advanced epileptic seizure detection from electroencephalography (EEG).
• Clinical adoption of AI for EEG analysis is hindered by a lack of model transparency and inability to explain seizure etiology.
• Existing AI models struggle to provide interpretable insights into the causes of seizures.

Purpose of the Study:

• To introduce SzXAI, a novel framework designed to enhance the reasoning capabilities of AI models for EEG-based seizure detection.
• To improve the transparency and interpretability of AI models used in epilepsy management.
• To bridge the gap between AI predictions and clinical understanding of seizure etiology.

Main Methods:

• SzXAI utilizes a contrastive training mechanism with cross-modality similarity layers.
• EEG encodings are aligned with textual concept embeddings derived from clinical notes using large language models (LLMs).
• An attention-weighted pooling mechanism is employed to detect underlying seizure and baseline etiologies.

Main Results:

• SzXAI was validated using 10-fold cross-validation on the Temple University Hospital dataset.
• The alignment-powered training mechanism significantly outperformed direct etiology prediction.
• Structured sentence generation provided human-readable insights, enhancing model interpretability.

Conclusions:

• SzXAI effectively improves the reliability of predicted seizure etiologies by enhancing AI reasoning.
• The framework boosts clinical trust and usability of AI tools in epilepsy management.
• SzXAI offers a promising platform for more transparent and interpretable AI-driven neurological diagnostics.