Generative modeling and augmentation of EEG signals using improved diffusion probabilistic models.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning for electroencephalography (EEG) signal processing faces data limitations.
• Data augmentation is crucial for improving model performance.
• Deep neural generative models offer potential for EEG data synthesis.

Purpose of the Study:

• Investigate enhanced diffusion probabilistic models (DPMs) for brain signal generation and EEG data augmentation.
• Evaluate the impact of implicit sampling and progressive distillation on generated data quality and inference time.
• Assess the effectiveness of DPM-augmented datasets in improving inter-subject EEG classification models.

Main Methods:

• Employed enhanced diffusion probabilistic models (DPMs) with implicit sampling and progressive distillation for EEG synthesis.
• Trained and evaluated four classification models on augmented datasets in an inter-subject setting.
• Analyzed generative metrics and statistical evaluations to assess signal quality and diversity.

Main Results:

• DPMs successfully generated visual evoked potentials and motor imagery EEG signals.
• Distilled, single-step DPMs synthesized high-quality EEG samples from public datasets.
• EEG classification model performance improved significantly with DPM-augmented data.
• Demonstrated high-fidelity data augmentation and improved EEG signal diversity using DPMs.

Conclusions:

• Diffusion probabilistic models show significant promise for EEG signal synthesis and data augmentation.
• DPMs offer an efficient and generalizable method for enhancing various EEG decoding tasks.
• A trade-off exists between data quality and sampling steps in single-step DPM generation.