Dissociating physiological ripples and epileptiform discharges with vision transformers

Area of Science:

• Neuroscience
• Computational Neuroscience
• Epilepsy Research

Background:

• Hippocampal ripples and interictal epileptiform discharges (IEDs) are distinct neural activity patterns.
• Existing automated ripple detectors often misidentify IEDs as false positives due to waveform similarities and artifacts.
• This overlap challenges independent study of ripples and IEDs, crucial for understanding epilepsy and cognitive function.

Purpose of the Study:

• To develop and validate a computer vision-based method for objective dissociation of hippocampal ripples and IEDs.
• To leverage time-frequency representations and deep learning models to improve detection accuracy.
• To assess the performance of vision transformers in distinguishing these neural events in human intracranial recordings.

Main Methods:

• Retrospective analysis of intracranial EEG recordings from 17 epilepsy patients.
• Application of a common ripple detection algorithm followed by spectrogram generation and k-means clustering for pseudo-labeling.
• Implementation of vision transformer models for spectrogram-based classification of ripple vs. IED candidates, validated against expert annotations.

Main Results:

• Spectrogram-based low-dimensional projections effectively separated canonical ripples and IEDs.
• Vision transformer models achieved high performance (AUC 0.970) in classifying IEDs vs. non-IEDs, generalizing well across patients (mean AUC 0.966).
• Attention maps indicated models focused on spectrotemporal features, including artifact patterns, to differentiate events.

Conclusions:

• The distinction between ripples and IEDs is better represented as a continuous gradient rather than a binary classification.
• Vision transformers demonstrate near-expert performance in dissociating ripples and IEDs by analyzing time-frequency spectrograms.
• This spectrotemporal analysis approach offers a powerful tool for advancing cognitive neurophysiology and optimizing biomarkers for epilepsy.