Addressing the Continuum Between Physiological Ripples and Interictal Discharges
Abstract number :
1.181
Submission category :
2. Translational Research / 2A. Human Studies
Year :
2024
Submission ID :
1012
Source :
www.aesnet.org
Presentation date :
12/7/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Da Zhang, PhD – University of California, San Francisco
Jonathan Kleen, MD, PhD – University of California, San Francisco
Rationale: Human intracranial recordings capture both normal and abnormal brain activity signatures since electrodes targeting potential seizure-generating areas are often implanted in brain tissue that has varying levels of residual function. Two of the most prominently studied intermittent neurophysiological events in the hippocampus are normal physiological ripples and abnormal interictal epileptiform discharges (IEDs), associated with normal or disrupted cognitive function, respectively. Waveform-based feature definitions help distinguish the two, yet this dichotomy ignores the potential variety of high frequency events in between, which may have gradations of functional significance for human memory and cognition investigations.
Methods: We studied human intracranial recordings from 65 hippocampal depth electrode sites among 18 patients with focal epilepsy, most with a seizure-onset zone/network involving the hippocampus. We implemented one of many prominent ripple detection algorithms from the literature, and omitted subsequent screening steps for IEDs which were therefore naturally included among the detections. We projected all detected events into low-dimensional space using uniform manifold approximation and projection. We adapted a vision transformer classification model with 5-fold cross-validation on canonical ripples vs. canonical IEDs to create a probability-based metric for estimating the degree of “epileptiformicity” of each candidate.
Results: We detected 40,031 ripple/IED candidates in the dataset. Projection into low (2-D) space illustrated a continuum of canonical normal ripples and canonical IEDs, which appeared at polar opposites of the projection, and far more intermediates in between. A hyperparameter search using K-means revealed that segmenting candidates into 7 clusters was roughly optimal. The binary transformer model trained on the two clusters representing canonical ripples (n=7825) and canonical IEDs (n=4006) achieved performance of 0.655-0.762 (range of 5 folds) for ROC-AUC, and between 0.769-0.815 for precision-recall (strongly significant based on 200 shuffled-label iterations). Intermediate detections (n=28,200) passed through the trained model revealed a positive skew of classification probabilities, with the mode closer to the ripple end of the spectrum, and only a trace of an additional IED-skewed mode (i.e. only minor bimodality).
Conclusions: The distinction between ripples and IEDs should be considered a continuous spectrum, and not dichotomous. This is likely due to the massive variety of sharpened waveform edges and pathological oscillations embedded with variable features of normal ripples, potentially undermining ripple- or IED-focused investigations using typical detectors. Probabilities (i.e. of being a ripple) for each detection candidate could instead serve as a quantitative surrogate for ripples (i.e. 0 to 1, not 0 or 1) as a new framework to address and embrace the vast gray zone between normal and abnormal high frequency detections. This work may fuel more comprehensive investigations in cognitive neurophysiology and biomarker optimization for closed-loop systems.
Funding: This work was supported by NINDS grants K23NS110920 (J.K.K.).
Translational Research