Characterizing physiological high-frequency oscillations using deep learning

Yipeng Zhang, Hoyoung Chung, Jacquline P. Ngo, Tonmoy Monsoor, Shaun A. Hussain, Joyce H. Matsumoto, Patricia D. Walshaw, Aria Fallah, Myung‐Shin Sim, Eishi Asano, Raman Sankar, Richard J. Staba, Jerome Engel, William Speier, Vwani Roychowdhury, Hiroki Nariai

Abstract

Intracranially-recorded interictal high-frequency oscillations (HFOs) are promising biomarkers for the epileptogenic zone. However, their occurrence in healthy brain tissue—physiological HFOs—limits their clinical utility. This study characterizes physiological HFOs using a weakly supervised deep learning (DL) model. Analyzing 63,379 HFOs from 18 children, we found that physiological HFOs (ecHFOs) exhibit lower amplitudes and a distinct "bell-shaped" time-frequency signature. Incorporating these DL-based classifications improved the prediction of postoperative seizure outcomes (AUC increased from 0.76 to 0.82), suggesting that automated filtering of physiological signals can better guide epilepsy surgery.

Introduction

In the surgical treatment of drug-resistant epilepsy, identifying the exact brain tissue responsible for seizures is paramount. High-frequency oscillations (HFOs), ranging from 80 to 500 Hz, have emerged as a critical spatial biomarker for the epileptogenic zone. Yet, the clinical application of HFOs is hindered by the fact that the healthy brain also generates these signals, particularly in "eloquent" cortex (EC) regions responsible for motor, sensory, and language functions.

Distinguishing "pathological" HFOs from "physiological" ones is a significant challenge for clinicians. Current manual or semi-automated methods are often insufficient to resolve this overlap, leading to potential over-resection of functional tissue or under-resection of the seizure-generating zone. This study presents a deep learning framework designed to automatically characterize and separate these signals.

Methods

The researchers employed a weakly supervised deep learning approach to analyze interictal intracranial EEG (iEEG) data from 18 pediatric patients with neocortical epilepsy.

Data Acquisition and Labeling

A total of 63,379 HFO events were recorded during 90-minute segments of sleep. To avoid the labor-intensive and subjective process of labeling individual events, the study used "channel-level" labels. Channels were identified as "eloquent cortex" (EC) through functional cortical mapping—direct electrical stimulation to identify critical functional areas. HFOs originating from these EC channels were labeled as physiological (ecHFOs).

Model Architecture

The team transformed one-dimensional EEG signals into three distinct 2D representations to utilize computer vision techniques:

  1. Time-Frequency Scalograms: Capturing power changes across frequencies.
  2. Raw EEG Tracings: Preserving the morphological shape of the waveform.
  3. Amplitude Coding Plots: Highlighting signal strength.

These images were processed using a ResNet-18 convolutional neural network. To ensure the model was learning relevant biological features rather than noise, the researchers used perturbation analysis, systematically altering the input data to observe changes in the model's classification confidence.

Results

Morphological Characteristics of Physiological HFOs

The DL model identified specific signatures that distinguish physiological HFOs:

  • The Bell-Shaped Template: ecHFOs typically presented a "bell-shaped" power distribution in time-frequency maps, characterized by lower amplitudes across the 80-500 Hz band around the HFO onset.
  • Absence of Spikes: A key differentiator was the presence of underlying EEG spikes. Only 22.9% of physiological HFOs were associated with spikes, whereas the vast majority of pathological HFOs (77.1%) occurred alongside these sharp discharges.
  • Amplitude and Frequency: While traditional features like duration showed overlap, the DL model successfully captured non-linear amplitude patterns that separated the two classes.

Clinical Validation and Outcome Prediction

The study validated the model's utility by correlating HFO resection with surgical outcomes in 14 patients. By calculating a "resection ratio" specifically for HFOs classified as pathological by the AI, the prediction of seizure freedom improved significantly. The Area Under the ROC Curve (AUC) rose from 0.76 (using uncorrected HFOs) to 0.82. When combined with the traditional Seizure Onset Zone (SOZ) clinical standard, the predictive accuracy reached an AUC of 0.93.

Discussion

This research demonstrates that deep learning can effectively distill the subtle morphological differences between healthy and pathological brain activity. The "bell-shaped" signature of physiological HFOs provides a new template for understanding normal brain oscillations in the high-frequency range.

While the results are promising, the study's small cohort (18 patients) necessitates further validation in larger, multi-center trials. Additionally, as clinical practice shifts toward stereo-EEG (SEEG), future iterations of the model will need to be tested on depth electrode recordings. Nevertheless, this DL-based classification offers a scalable, objective method to refine the use of HFOs in surgical planning, potentially improving outcomes for children with refractory epilepsy.

Summary updated

Abstract

<i>Objective.</i>Intracranially-recorded interictal high-frequency oscillations (HFOs) have been proposed as a promising spatial biomarker of the epileptogenic zone. However, HFOs can also be recorded in the healthy brain regions, which complicates the interpretation of HFOs. The present study aimed to characterize salient features of physiological HFOs using deep learning (DL).<i>Approach.</i>We studied children with neocortical epilepsy who underwent intracranial strip/grid evaluation. Time-series EEG data were transformed into DL training inputs. The eloquent cortex (EC) was defined by functional cortical mapping and used as a DL label. Morphological characteristics of HFOs obtained from EC (ecHFOs) were distilled and interpreted through a novel weakly supervised DL model.<i>Main results.</i>A total of 63 379 interictal intracranially-recorded HFOs from 18 children were analyzed. The ecHFOs had lower amplitude throughout the 80-500 Hz frequency band around the HFO onset and also had a lower signal amplitude in the low frequency band throughout a one-second time window than non-ecHFOs, resembling a bell-shaped template in the time-frequency map. A minority of ecHFOs were HFOs with spikes (22.9%). Such morphological characteristics were confirmed to influence DL model prediction via perturbation analyses. Using the resection ratio (removed HFOs/detected HFOs) of non-ecHFOs, the prediction of postoperative seizure outcomes improved compared to using uncorrected HFOs (area under the ROC curve of 0.82, increased from 0.76).<i>Significance.</i>We characterized salient features of physiological HFOs using a DL algorithm. Our results suggested that this DL-based HFO classification, once trained, might help separate physiological from pathological HFOs, and efficiently guide surgical resection using HFOs.
BibTeX
@article{zhang2022characterizing,
  author = {Yipeng Zhang and Hoyoung Chung and Jacquline P. Ngo and Tonmoy Monsoor and Shaun A. Hussain and Joyce H. Matsumoto and Patricia D. Walshaw and Aria Fallah and Myung‐Shin Sim and Eishi Asano and Raman Sankar and Richard J. Staba and Jerome Engel and William Speier and Vwani Roychowdhury and Hiroki Nariai},
  title = {{Characterizing physiological high-frequency oscillations using deep learning}},
  journal = {Journal of Neural Engineering},
  year = {2022},
  doi = {10.1088/1741-2552/aca4fa},
}
Figure