Refining epileptogenic high-frequency oscillations using deep learning: a reverse engineering approach

Yipeng Zhang, Qiujing Lu, Tonmoy Monsoor, Shaun A. Hussain, Joe X Qiao, Noriko Salamon, Aria Fallah, Myung‐Shin Sim, Eishi Asano, Raman Sankar, Richard J. Staba, Jerome Engel, William Speier, Vwani Roychowdhury, Hiroki NariaibioRxiv 2021 · DOI

TL;DR

This paper introduces a novel deep learning (DL) approach to automatically identify epileptogenic high-frequency oscillations (eHFOs) in EEG data, overcoming the limitations of manual analysis. By using a weakly supervised method trained on clinical outcomes, the DL model not only accurately classifies HFOs but also uncovers key features of eHFOs, demonstrating its potential to improve epilepsy surgery planning.

What Problem Does This Paper Address?

Epileptogenic high-frequency oscillations (HFOs) are promising biomarkers for pinpointing the seizure-generating regions in the brain (the epileptogenic zone or EZ). However, manually identifying and classifying these HFOs is a time-consuming process with significant variability between experts (poor inter-rater reliability). Crucially, there's no reliable method to distinguish HFOs originating from the EZ (epileptogenic HFOs or eHFOs) from those originating from healthy brain regions (non-epileptogenic HFOs or non-eHFOs). This distinction is vital for successful epilepsy surgery, as leaving eHFOs behind can lead to continued seizures.

Methods

The researchers developed a two-step deep learning framework. First, a supervised DL model was trained to mimic human experts in classifying HFOs into three categories: HFOs with spikes, HFOs without spikes, and artifacts. This step achieved high accuracy in replicating expert annotations.

The second, more innovative step, focused on discovering eHFOs using a weakly supervised approach. Instead of relying on expert labels for eHFOs, the model was trained using clinical outcomes (seizure freedom after surgery) and the location of electrodes (resected vs. preserved brain regions). The assumption was that HFOs in resected regions of seizure-free patients are likely epileptogenic, while those in preserved regions are not. This "weak labeling" was then refined by the DL model's "purification power" to automatically relabel HFOs and identify eHFOs.

To understand how the model made its decisions, a "reverse engineering" approach was employed. This involved perturbing the input data (e.g., altering time-frequency plots or amplitude-coding) and observing the impact on the model's output probability. This helped identify the salient features the DL model learned to associate with eHFOs.

Datasets

The study utilized intracranial EEG (iEEG) data from 19 children with medication-resistant epilepsy who underwent surgery.

  • Training Data:
    • 12,958 annotated HFO events from 19 patients for artifact detection and HFO with/without spike classification.
  • 84,602 HFO events from 9 seizure-free patients for training the weakly supervised eHFO discovery model.
  • Inference/Validation Data:
    • 90-minute EEG segments from 15 patients who underwent resection were used for applying the trained DL models and mapping eHFOs.
  • 14 patients with known 24-month post-operative seizure outcomes were used for predicting surgical success based on HFO resection ratios.

No new benchmarks were introduced; the study focused on applying DL to existing iEEG data and clinical outcomes.

Key Insights and Ideas

  • DL as an Expert Mimic: Deep learning can effectively replicate the nuanced classifications of human experts in identifying HFOs and artifacts, potentially reducing the burden on clinicians.
  • Weak Supervision for Discovery: Leveraging clinical outcomes (like seizure freedom) and anatomical information (resected vs. preserved tissue) as "weak labels" is a powerful way to discover novel biomarkers like eHFOs, bypassing the need for perfect expert annotations.
  • Reverse Engineering for Interpretability: By systematically perturbing model inputs, researchers can gain insights into the features that drive DL model predictions, making these "black box" models more understandable and trustworthy for clinical applications.
  • eHFOs as Predictive Biomarkers: The resection ratio of eHFOs (the proportion of detected eHFOs found in resected tissue) is a strong predictor of post-operative seizure freedom, outperforming general HFO detection.
  • Salient eHFO Features: The DL model learned to associate eHFOs with specific time-frequency patterns (an "inverted T-shaped" signature) and the presence of spike-like activity, aligning with existing expert knowledge.

Key Results

  • Expert Emulation: The DL model achieved high accuracy: 96.3% on artifact detection (F1 score 96.8%) and 86.5% on classifying HFOs with spikes (F1 score 80.8%) in a patient-wise cross-validation.
  • eHFO Discovery: The weakly supervised model identified eHFOs that were clinically relevant. A significant majority (71.1%) of DL-discovered eHFOs were indeed HFOs with spikes, as identified by experts.
  • Predictive Power of eHFOs: The resection ratio of eHFOs showed a significant positive correlation with post-operative seizure freedom (AUC = 0.87, p = 0.01), outperforming the resection ratio of all detected HFOs (AUC = 0.76, p = 0.06).
  • Spatial Localization: eHFOs were found to cluster around the seizure onset zone (SOZ), while non-eHFOs were more diffusely distributed.
  • Feature Identification: Perturbation analysis revealed that introducing signals in the "inverted T-shaped" frequency band and spike-like patterns in the time domain significantly increased the model's confidence in classifying an HFO as epileptogenic.

Limitations and Future Work

  • Dataset Size: The study included a relatively small cohort of 19 patients, with only 14 having known post-operative outcomes. The authors suggest that including more patients and data from diverse institutions is crucial for further refining the DL algorithm and improving its generalizability.
  • Pediatric Focus: All data were from pediatric patients, and the authors acknowledge that HFO morphology might vary with age and epilepsy pathology, potentially requiring different training strategies for adult populations.
  • Channel Resection Status Subjectivity: While the study used resected vs. preserved channel status as a weak label, this determination was made by human experts reviewing images, which could introduce some subjectivity. Future work could explore more automated and objective methods for determining channel resection status.
  • Future Directions: The authors plan to build models using data from over 100 patients and explore the discovery of new salient characteristics of eHFOs as DL models become more robust. They also aim to refine the method to distinguish epileptogenic HFOs more precisely, ultimately aiming to guide resection margins in clinical trials to improve post-operative seizure outcomes.

Summary updated

Abstract

ABSTRACT Intracranially-recorded interictal high-frequency oscillations (HFOs) have been proposed as a promising spatial biomarker of the epileptogenic zone. However, visual verification of HFOs is time-consuming and exhibits poor inter-rater reliability. Furthermore, no method is currently available to distinguish HFOs generated from the epileptogenic zone (epileptogenic HFOs: eHFOs) from those generated from other areas (non-epileptogenic HFOs: non-eHFOs). To address these issues, we constructed a deep learning (DL)-based algorithm using HFO events from chronic intracranial electroencephalogram (iEEG) data via subdural grids from 19 children with medication-resistant neocortical epilepsy to: 1) replicate human expert annotation of artifacts and HFOs with or without spikes, and 2) discover eHFOs by designing a novel weakly supervised model (HFOs from the resected brain regions are initially labeled as eHFOs, and those from the preserved brain regions as non-eHFOs). The “purification power” of DL is then used to automatically relabel the HFOs to distill eHFOs. Using 12,958 annotated HFO events from 19 patients, the model achieved 96.3% accuracy on artifact detection (F1 score = 96.8%) and 86.5% accuracy on classifying HFOs with or without spikes (F1 score = 80.8%) using patient-wise cross-validation. Based on the DL-based algorithm trained from 84,602 HFO events from nine patients who achieved seizure-freedom after resection, the majority of such DL-discovered eHFOs were found to be HFOs with spikes (78.6%, p < 0.001). While the resection ratio of detected HFOs (number of resected HFOs/number of detected HFOs) did not correlate significantly with post-operative seizure freedom (the area under the curve [AUC]=0.76, p=0.06), the resection ratio of eHFOs positively correlated with post-operative seizure freedom (AUC=0.87, p=0.01). We discovered that the eHFOs had a higher signal intensity associated with ripple (80-250 Hz) and fast ripple (250-500 Hz) bands at the HFO onset and with a lower frequency band throughout the event time window (the inverted T-shaped), compared to non-eHFOs. We then designed perturbations on the input of the trained model for non-eHFOs to determine the model’s decision-making logic. The model probability significantly increased towards eHFOs by the artificial introduction of signals in the inverted T-shaped frequency bands (mean probability increase: 0.285, p < 0.001), and by the artificial insertion of spike-like signals into the time domain (mean probability increase: 0.452, p < 0.001). With this DL-based framework, we reliably replicated HFO classification tasks by human experts. Using a reverse engineering technique, we distinguished eHFOs from others and identified salient features of eHFOs that aligned with current knowledge. Graphical abstract
BibTeX
@misc{zhang2021refining,
  author = {Yipeng Zhang and Qiujing Lu and Tonmoy Monsoor and Shaun A. Hussain and Joe X Qiao and Noriko Salamon 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 = {{Refining epileptogenic high-frequency oscillations using deep learning: a reverse engineering approach}},
  year = {2021},
  doi = {10.1101/2021.08.31.458385},
  howpublished = {bioRxiv},
}
Figure