bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2025, Volume and Issue: unknown
Published: April 18, 2025
Abstract Two frequently studied bursts of neural activity in the hippocampus are normal physiological ripples and abnormal interictal epileptiform discharges (IEDs). While they different waveforms, IEDs notoriously picked up as false positives when using typical automated detectors which prone to sharp edge artifacts. This has created challenges for studying independently. We leveraged recent advances computer vision on time-frequency feature representations enable more comprehensive objective dissociation these phenomena. retrospectively evaluated human intracranial recordings from 46 hippocampal depth electrode sites among 17 patients with focal epilepsy, majority whom had a seizure-onset zone/network involving hippocampus. implemented common ripple detection algorithm broadband spectrograms all detected “ripple candidates” were projected into low-dimensional space. segmented them k-means infer pseudo-labels probable IEDs. Independently, expert IED labels manually annotated comparison. State-of-the-art transformer models individual approach vs. an image classification problem. 31,847 ripple/IED candidates, median 3.9% per patient (range: 0-47.2%) based label overlap. Low-dimensional projection separated canonical better than raw or ripple-filtered waveforms. Canonical candidates emerged at opposite poles continuous landscape intermediates between. A binary model trained expert-labeled non-IED candidate 5-fold cross-validation showed mean area under curve (AUC) 0.970 precision-recall 0.694, both significantly above chance. To evaluate generalizability, we leave-one-patient-out approach, training testing data demonstrated near-expert performance (mean AUC 0.966 across patients, range 0.892-0.997). Transformer-derived attention maps revealed that tuned triangle-like artifact spatial features spectrograms. Model-derived probabilities (i.e. being IED) transitions between IEDs, opposed clustering. The delineation appears best represented gradient not binary) due overlapping sharpened and/or high frequency pathophysiological features. Vision transformers nevertheless perform virtually levels dissociating phenomena by leveraging enabled Such tools applied spectrotemporal may augment investigations cognitive neurophysiology signal biomarker optimization closed-loop applications.
Language: Английский