Evaluating resective surgery targets in epilepsy patients: A comparison of quantitative EEG methods

Michael Müller, Kaspar Schindler, Marc Goodfellow, Claudio Pollo, Christian Rummel, Andreas SteimerJournal of Neuroscience Methods 2018 · DOI

TL;DR

This paper compares two distinct quantitative EEG (qEEG) methods for identifying optimal targets for epilepsy surgery. Both methods show promise in predicting surgical outcomes, and while they operate on different principles, their assessments of hypothetical resections show a significant positive correlation, suggesting a consensus that can aid clinical decision-making.

What Problem Does This Paper Address?

Epilepsy surgery aims to remove brain tissue responsible for seizures, but precisely identifying this "epileptogenic zone" (EZ) is challenging. The seizure onset zone (SOZ), often used as a proxy, is typically identified visually, which is time-consuming and prone to variability. Quantitative EEG (qEEG) methods offer a data-driven approach to assist in this process, but with multiple methods available, it's unclear how they compare and if combining them offers benefits. This paper addresses the critical need to evaluate and compare different qEEG approaches, particularly those with fundamentally different underlying principles, to understand their agreement and potential for clinical integration. The lack of a definitive "ground truth" for hypothetical resections makes direct accuracy assessment difficult, necessitating comparative studies like this one.

Methods

The study compares two qEEG methods:

  • Distributional Soft Clustering (SC): This method uses machine learning to build a probabilistic model of brain states from peri-ictal iEEG data. It can then predict how likely a seizure state is under simulated resections, effectively quantifying the "seizure suppressing efficiency" of a hypothetical resection.
  • Multivariate Nonlinear Interrelation-based Functional Networks (FN): This approach models the brain as a network where channels are nodes and edges represent nonlinear interrelations (quantified by mutual information). A "node strength" is calculated for each channel, and the performance of a resection is assessed by the collective node strength of the resected channels relative to the total network strength.

Both methods were applied to the same dataset of patients undergoing epilepsy surgery. The study evaluated their ability to distinguish between patients with good (Engel class I) and poor (Engel class IV) post-surgical outcomes, compared their rankings of patients, and investigated the impact of combining their predictions using AND and OR conjunctions. They also explored how the methods' assessments of hypothetical resections relate to the overlap with the actual resected tissue.

Datasets

The study utilized intracranial EEG (iEEG) recordings from 20 patients who underwent epilepsy surgery at the Inselspital Bern. These patients were selected based on their post-surgical outcome (Engel class I or IV) and the availability of information on resected brain tissue and electrode placement. The dataset is not new, but it serves as a benchmark for comparing the two qEEG methods.

Key Insights and Ideas

  • Consensus Despite Different Approaches: The most striking finding is the significant positive correlation between the patient rankings generated by the SC and FN methods. This suggests that despite their fundamentally different underlying principles (machine learning vs. network theory), they often agree on which patients are likely to benefit from surgery.
  • Overlap Matters: Both methods' assessments of hypothetical resections were influenced by the overlap with the actual resected tissue, particularly for patients with good outcomes. This indicates that the methods are sensitive to the anatomical regions that were historically targeted.
  • No Clear Benefit from Simple Combination: Combining the binary classifiers of the two methods using AND or OR conjunctions did not consistently improve classification performance over individual methods. This suggests that while they agree, their agreement doesn't necessarily lead to a simple additive benefit in binary classification.
  • The "Ground Truth" Problem: The paper highlights the inherent difficulty in validating these methods due to the lack of a definitive ground truth for hypothetical resections. We can only observe the outcome of the actual resection, making it hard to definitively assess the sensitivity and specificity of predictions for other potential resections.

Key Results

  • Distinguishing Outcomes: Both the SC and FN methods were individually capable of distinguishing between class I and class IV patients based on their actual resections (p < 0.0004 for both).
  • Patient Ranking Correlation: The Spearman rank correlation between the patient rankings from the SC and FN methods was significant (ρ = 0.60, p = 0.0027).
  • ROC AUC: The Area Under the ROC Curve (AUC) for the SC method was 0.85, and for the FN method, it was 0.86, indicating good individual patient-level classification performance.
  • Overlap Dependence: For class I patients, virtual resections with greater overlap to the actual resection generally received higher ratings from both methods. This effect was less pronounced or absent in class IV patients.
  • Conditional Independence: Partial correlation analysis suggested that the methods' ratings are conditionally independent given the overlap with the actual resection, implying their agreement isn't solely driven by this overlap.

Limitations and Future Work

  • Small Sample Size: The study acknowledges the limitation of a small sample size (n=20), which restricts the statistical power and generalizability of the findings. Bootstrapping was used to mitigate this, but larger datasets are needed.
  • Ground Truth Absence: As mentioned, the lack of a true ground truth for hypothetical resections remains a significant challenge for objective validation.
  • Misclassifications: Both methods misclassified at least one patient, highlighting that neither is perfect. Patient 15 was a notable case where both methods predicted a good outcome, but the actual surgery was ineffective.
  • Future Work: The authors suggest further studies with larger, unselected datasets to confirm these findings and establish the clinical utility of these methods. They also propose exploring more sophisticated combinations of methods and investigating the underlying signal features that drive their agreement and disagreement. The potential for these methods to objectify and simplify the presurgical evaluation process is a key future direction.

Summary updated

Abstract

Although conceptually completely independent, there is a relation between the predictions obtained from both methods. Their broad consensus supports their application in clinical practice to provide physicians additional information in the process of presurgical evaluation.
BibTeX
@article{muller2018evaluating,
  author = {Michael Müller and Kaspar Schindler and Marc Goodfellow and Claudio Pollo and Christian Rummel and Andreas Steimer},
  title = {{Evaluating resective surgery targets in epilepsy patients: A comparison of quantitative EEG methods}},
  journal = {Journal of Neuroscience Methods},
  year = {2018},
  doi = {10.1016/j.jneumeth.2018.04.021},
}
Figure