SEEG-Net: An explainable and deep learning-based cross-subject pathological activity detection method for...

Area of Science:

• Artificial Intelligence in Medicine
• Neuroscience
• Medical Imaging Analysis

Background:

• Accurate analysis of invasive stereoelectroencephalography (SEEG) is crucial for drug-resistant epilepsy (DRE) preoperative evaluation.
• Existing AI models for SEEG analysis face challenges like sample imbalance, cross-subject domain shift, and poor interpretability in real-world clinical DRE scenarios.
• There is a need for advanced AI solutions to improve the sensitivity and reliability of SEEG pathological activity detection.

Purpose of the Study:

• To propose an AI model, SEEG-Net, designed to overcome the limitations of current methods in SEEG pathological activity detection for DRE.
• To achieve high-sensitivity detection of pathological activities in SEEG data from real clinical datasets.
• To enhance the interpretability and flexibility of AI models in clinical epilepsy research.

Main Methods:

• Development of SEEG-Net, incorporating a multiscale convolutional neural network (MSCNN) to capture multi-frequency domain, local, and global SEEG features.
• Introduction of a novel focal domain generalization loss (FDG-loss) function to improve target sample weighting and learn domain-consistent features.
• Utilizing Grad-CAM++ for multi-perspective explanations of SEEG-Net, enhancing model interpretability and flexibility.

Main Results:

• The SEEG-Net model demonstrated state-of-the-art performance on a public multicenter SEEG dataset and a private clinical SEEG dataset.
• Achieved the highest sensitivity in cross-subject evaluations, effectively addressing known issues in SEEG data analysis.
• Established a consistent SEEG processing and database construction workflow aligned with real-world clinical practices.

Conclusions:

• The developed SEEG-Net significantly increases the sensitivity of SEEG pathological activity detection for drug-resistant epilepsy.
• The study successfully addressed key challenges in applying AI assistance to clinical DRE management.
• Established a crucial link between advanced AI algorithm applications and practical clinical epilepsy care.