A Composable Channel-Adaptive Architecture for Seizure Classification.

Area of Science:

• Deep learning applications in neuroscience and healthcare.
• Multivariate time-series analysis for biomedical signals.

Background:

• Deep learning models excel with multivariate time-series data, common in healthcare.
• Intracranial electroencephalography (iEEG) is crucial for various medical tasks.
• Current iEEG models struggle with the variable number of channels in personalized clinical setups.

Purpose of the Study:

• To introduce a channel-adaptive (CA) architecture for iEEG analysis.
• To enable deep learning models to process multivariate signals with arbitrary channel counts.
• To improve the efficiency and performance of seizure detection models.

Main Methods:

• Developed a novel channel-adaptive (CA) architecture for deep learning models.
• Pre-trained CA models on large, diverse iEEG datasets.
• Fine-tuned CA models on subject-specific data, requiring less data and time.
• Evaluated CA models (CA-EEGWaveNet, CA-EEGNet) on seizure detection tasks using short-term and long-term datasets.

Main Results:

• CA models demonstrated seamless functionality with varying numbers of iEEG channels.
• Fine-tuning CA models required significantly less data and time (1/5) compared to existing methods.
• CA-EEGWaveNet achieved a median F1-score of 0.78, outperforming the baseline (0.76).
• CA-EEGNet achieved a median F1-score of 0.79, surpassing its baseline (0.74).

Conclusions:

• The channel-adaptive architecture serves as a versatile, drop-in replacement for current seizure classification models.
• CA architecture offers improved performance and characteristics for iEEG analysis.
• This approach enhances the adaptability and efficiency of deep learning in personalized neurological applications.