Discovery of Neurophysiological Characteristics of Pathological High-frequency Oscillations in Epilepsy with an Explainable Deep Generative Model
Abstract number :
1.302
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2024
Submission ID :
1142
Source :
www.aesnet.org
Presentation date :
12/7/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Yipeng Zhang, Dr. – UCLA
Atsuro Daida, MD,PhD – UCLA Mattel Children's Hospital
Lawrence Liu, BS – University of California, Los Angeles
Naoto Kuroda, MD – Wayne State University
Yuanyi Ding, MS – University of California, Los Angeles
Shingo Oana, MD, PhD – University of California, Los Angeles
Tonmoy Monsoor, PhD – University of California, Los Angeles
Shaun Hussain, MD, MS – UCLA Mattel Children's Hospital, David Geffen School of Medicine
Joe Qiao, PhD – University of California, Los Angeles
Noriko Salamon, MD,PhD – UCLA Mattel Children's Hospital
Aria Fallah, MD, MS – UCLA Mattel Children's Hospital
Myung Shin Sim, PhD – University of California, Los Angeles
Raman Sankar, MD, PhD – University of California, Los Angeles
Richard Staba, PhD – University of California, Los Angeles
Jerome Engel Jr., MD, PhD – University of California, Los Angeles
Eishi Asano, MD/PhD – Wayne State University
Vwani Roychowdhury, PhD – UCLA
Hiroki Nariai, MD, PhD, MS – UCLA Mattel Children's Hospital
Rationale: Interictal high-frequency oscillations (HFOs) is a promising neurophysiological biomarker of the epileptogenic zone (EZ). HFOs with similar frequency ranges, however, also exist in healthy brain regions, and objective definitions to distinguish between pathological and physiological HFOs have remained elusive, impeding HFOs’ clinical applications. We hypothesize that such distinction between pathological and physiological HFOs is purely in their morphology and can be learned in a data-driven manner through self-supervised learning using a deep generative model, variational autoencoder (VAE).
Methods: We studied a retrospective cohort of 185 epilepsy patients who underwent intracranial monitoring with grid/strip or stereotactic EEG electrodes. We sampled 18,265 brain contacts across 34 regions of interest. After identifying 686,410 HFOs using an automated detector, each HFO event's EEG time-series data was transformed into time-frequency analysis imaging data. This data served as the input for the VAE. During training, the model was tasked with reconstructing the input time-frequency plot, ensuring the latent space followed a Gaussian distribution. The cluster with high reconstruction loss was defined as the cluster of artifacts (mArtifact). To let the model allocate either a pathological (morphologically determined pathological HFOs: mpHFO) or physiological (non-mpHFO) label to each cluster, we employed a minimalistic use of clinical data. The cluster with a higher resection percentage in seizure-free patients after resection was deemed pathological. The Gaussian Mixture Models were then used to assign predictions, mArtifact, pathological or physiological, on all HFOs' latent codes from test set patients.
Results: mpHFOs predominantly were associated with spikes and originated within the seizure onset zones (SOZ) (Fig. a-c). These mpHFOs possessed consistent characteristics: high signal intensity within the HFO band (≥ 80 Hz) at detection and in the sub-HFO band (10-80 Hz) surrounding the detection, resembling a "hanging bell" shape in a time-frequency plot (Fig. d-e). Their sub-HFO band peak frequency was 23 Hz (Fig. f-g). Such non-HFO band EEG signals associated with mpHFOs seemed to emerge as spikes in time series data (Fig. h-j), contrasting from non-mpHFOs (Fig. k-m). With the random forest model trained through five-fold cross-validation, the resection ratio of mpHFOs exhibited better predictive performance on postoperative seizure outcomes (F1 = 0.74) compared to using unclassified HFOs (F1 = 0.69, p-value < 0.01). The combination of demographic data and mpHFO resection ratio (F1 = 0.77) outperformed traditional clinical predictions based on demographic information and SOZ resection status (F1 = 0.73, p-value < 0.01). Furthermore, a comprehensive model including all features, including demographic data, SOZ resection, and mpHFO resection ratio, demonstrated an excellent prediction performance (F1 = 0.81).
Conclusions: Our explainable deep generative model can distinguish pathological HFOs in a self-supervised manner and maximize the utility of HFOs to help delineate the EZ.
Funding: The National Institute of Health, K23NS128318
Neurophysiology