Authors :
Presenting Author: Jacqueline Ngo, MD – UCLA
Yipeng Zhang, MS – Electrical and Computer Engineering – UCLA; Lawrence Liu, MS – Electrical and Computer Engineering – UCLA; Tonmoy Monsoor, MS – Electrical and Computer Engineering – UCLA; Atsuro Daida, MD, PhD – Pediatrics – UCLA; Shingo Oana, MD, PhD – Pediatrics – UCLA; Shaun Hussain, MD, MS – Pediatrics – UCLA; Raman Sankar, MD, PhD – Pediatrics – UCLA; Aria Fallah, MD, MS – Neurosurgery – UCLA; Richard Staba, PhD – Neurology – UCLA; Jerome Engel, MD, PhD – Neurology – UCLA; William Speier, PhD – Radiological Sciences and Bioengineering – UCLA; Vwani Roychowdhury, PhD – Electrical and Computer Engineering – UCLA; Hiroki Nariai, MD, PhD, MS – Assistant Professor, Pediatrics, UCLA
Rationale:
Interictal high-frequency oscillations (HFOs) are considered one of the promising spatial neurophysiological biomarkers of the epileptogenic zone. However, the task of distinguishing pathological HFOs from physiological ones presents a significant challenge, yet it's crucial for their clinical application. We hypothesize that the distinctive morphological features of pathological HFOs can be discerned from physiological HFOs using an unsupervised learning approach, negating the need for pre-assigned training labels.
Methods:
We used chronic intracranial electroencephalogram (iEEG) data through subdural grids from 18 pediatric patients with medication-resistant neocortical epilepsy. After identifying 92,860 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 deep learning model, specifically a variational autoencoder (VAE). During training, the model was tasked with reconstructing the input time-frequency plot, ensuring the latent space followed a Gaussian distribution. This unsupervised approach didn't require labels indicating whether an event was pathological. Post-training, the HFO events' latent codes, stratified from all training patients, were clustered by the Gaussian Mixture Model (GMM) with K=2. The cluster with a higher association with resection in post-surgical seizure-free patients was deemed pathological. The GMM model was then used to assign predictions, pathological or physiological, on all HFOs' latent codes from test set patients.
Results:
The effectiveness of our unsupervised method was gauged through a stringent patient-wise, five fold cross-validation. We projected randomly selected HFOs' latent codes into a two-dimensional (2D) space, comparing the pathological predictions from the VAE model with HFO-with-spike, as such HFOs are generally more pathological. Our analysis reveals that the pathological prediction from the VAE closely aligns with HFO-with-spike (Figure 1). Moreover, pathological HFOs, as predicted by our VAE model, established a pattern in the time-frequency plot (Figure 2). This pattern closely resembled the structure of an inverted T-shaped template, exhibiting characteristics akin to the ones we identified in our prior research (Zhang et al. Brain Commun. 2021. PMID: 35169696). Using the resection ratio of pathological HFOs, as predicted by the VAE model, to forecast postoperative seizure outcomes resulted in an AUC of 0.91 (p = 1.017e-5), signifying a significant improvement compared to the AUC of 0.82 (p = 3.925e-5) obtained using the resection ratio of unclassified HFOs. Additionally, the VAE model outperformed the AUC of 0.89 (p = 1.307e-5) achieved using the resection ratio of HFOs with spikes.
Conclusions:
We have demonstrated the ability to classify pathological HFOs using unsupervised machine learning with VAE, eliminating the need for any labeling. This approach could significantly enhance the clinical utility of pathological HFOs, particularly in delineating the epileptogenic zone during epilepsy surgery.
Funding:
The National Institute of Health, K23NS128318