Unsupervised EEG-Based Seizure Anomaly Detection with Denoising Diffusion Probabilistic Models.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate seizure detection methods often require extensive labeled data.
• Unsupervised approaches are needed to overcome data limitations in seizure detection.

Purpose of the Study:

• To develop a novel unsupervised seizure anomaly detection method named SAnoDDPM.
• To leverage denoising diffusion probabilistic models (DDPM) for efficient and accurate seizure detection.

Main Methods:

• SAnoDDPM employs a pipeline with variable lower bounds on Markov chains for anomaly identification.
• 2D spectrograms are encoded into vector-quantized representations for enhanced DDPM performance.
• The model is trained on normal data and then used to convert anomalous (seizure) data back to normal patterns.

Main Results:

• SAnoDDPM demonstrates superior performance on CHB-MIT and TUH datasets compared to existing methods.
• The method significantly reduces inference time, making it suitable for clinical deployment.
• This is the first reported application of DDPM for seizure anomaly detection.

Conclusions:

• SAnoDDPM offers an effective unsupervised approach for seizure detection, addressing the need for large labeled datasets.
• The method's efficiency and accuracy enhance the practicality of seizure detection algorithms in clinical settings.