TL;DR
This paper introduces a novel method to quantify the deviation of intracranial electrocorticography (iEEG) signals from a patient-specific normative baseline, inspired by statistical parametric mapping (SPM). This "MI z-score" significantly improves the prediction of surgical success in epilepsy patients, demonstrating its potential as a valuable biomarker for presurgical evaluation.
What Problem Does This Paper Address?
Epilepsy surgery aims to remove the seizure onset zone (SOZ) to achieve seizure freedom. However, accurately identifying the SOZ, especially in drug-resistant cases, remains a significant challenge. While high-frequency activity (HFA) and interictal spikes are known indicators, their localization can be imprecise, and they can also occur in non-epileptic regions. Previous work has shown that the "modulation index" (MI), a measure of phase-amplitude coupling between HFA (>150 Hz) and slow waves (3-4 Hz), can be a useful biomarker. However, the inherent variability of MI across different brain regions and individuals necessitates a way to quantify how much a specific recording site's MI deviates from a "normal" or expected value for that region. This paper tackles this by adapting the concept of statistical parametric mapping (SPM) to iEEG data, creating a patient-specific normative baseline for MI and using this deviation to improve SOZ localization and surgical outcome prediction.
Methods
The core innovation is the development of a "MI z-score." This involves:
- Normative MI Atlas Generation: Using iEEG data from 47 patients with electrodes across all four lobes, a normative atlas of slow-wave sleep MI was created. This atlas represents the mean and standard deviation of MI across non-epileptic channels for each vertex on a standard brain surface.
- MI z-score Calculation: For each individual patient, the MI was calculated at each electrode site. Then, the "MI z-score" was computed for each site, representing how many standard deviations its MI value was from the normative mean for that specific brain region. This effectively quantifies the statistical deviation of MI from the expected value.
- Seizure Outcome Prediction: A multivariate logistic regression model was built to predict surgical success (defined as ILAE Class 1 outcome). This model incorporated various clinical, neuroimaging, and iEEG variables, including the newly derived "MI z-score" (specifically, a "subtraction MI z-score" representing the difference between resected and preserved tissue MI z-scores).
- Model Evaluation: The performance of the full model (including MI z-score) was compared against models that excluded MI z-score or SOZ information. Receiver operating characteristic (ROC) analysis was used to assess classification accuracy (AUC, sensitivity, specificity).
- Cross-Validation: A leave-one-out cross-validation approach was employed to ensure the robustness of the model's predictive power.
Datasets
The study utilized retrospective iEEG data from 123 patients with drug-resistant focal epilepsy who underwent resective epilepsy surgery.
- Normative Atlas Development: A subset of 47 patients with electrodes covering all four lobes was used to generate the normative MI atlas.
- Model Validation: The full cohort of 123 patients was used for training and validating the predictive models.
The dataset is characterized by iEEG recordings with a sampling rate of 1000 Hz, collected between January 2007 and October 2016.
Key Insights and Ideas
- SPM for iEEG: The paper successfully adapts the concept of statistical parametric mapping, traditionally used in non-invasive neuroimaging, to quantify deviations in invasive iEEG signals. This opens up new avenues for analyzing iEEG data with a focus on statistical significance relative to a normative baseline.
- MI z-score as a Biomarker: The "MI z-score" is not just a descriptive measure but a powerful predictive biomarker. Its statistical deviation from the normative mean appears to be a more robust indicator of epileptogenicity than raw MI values alone.
- Localizing the "Irritative Zone": The study suggests that regions with significantly higher MI z-scores (deviating from the norm) are more likely to be part of the irritative zone, which often overlaps with the SOZ.
- Synergy of Predictors: The full model, incorporating clinical, SOZ, and MI z-score variables, significantly outperforms models that omit the MI z-score, highlighting its unique contribution to predicting surgical outcomes.
- Feasibility of Invasive SPM: The research demonstrates that creating a normative baseline and calculating statistical deviations is technically feasible even with invasive iEEG recordings, overcoming the limitations of traditional SPM which relies on large control groups.
Key Results
- SOZ vs. Non-SOZ MI z-score: In both patient groups analyzed (47 and 64 patients), the SOZ electrodes consistently showed significantly higher "MI z-scores" compared to non-SOZ electrodes (grand-medians of 0.96 vs. 0.13 and 1.28 vs. 0.52, respectively, both with p < 0.001).
- Predictive Power of the Full Model: The full multivariate logistic regression model, including "subtraction MI z-score," achieved an AUC of 0.85 for classifying surgical success. With a sensitivity of 0.86, it yielded a specificity of 0.76.
- Impact of MI z-score: Excluding the "MI z-score" from the model significantly reduced its predictive accuracy, lowering the specificity to 0.48 while maintaining the same sensitivity (0.86).
- Leave-One-Out Validation: The leave-one-out cross-validation confirmed the robustness of the full model, with an AUC of 0.77. Models excluding MI or SOZ showed substantially lower AUCs (0.61 and 0.66, respectively) in this cross-validation.
- Clinical Significance: A larger "subtraction MI z-score" (indicating removal of high MI z-score regions and preservation of low MI z-score regions) was associated with an increased chance of surgical success (OR: 1.45, p=0.01).
Limitations and Future Work
- Retrospective Study: The study is retrospective, which inherently limits causal inference and may be subject to biases in data collection and interpretation.
- Single Institution: The data was collected from a single institution, potentially limiting the generalizability of findings to other clinical settings.
- Age Range: Children younger than 4 years were excluded due to potential issues with FreeSurfer registration. The authors acknowledge that normative MI values might differ in this age group.
- Brain Malformations: The method was not applicable to patients with major brain malformations (e.g., large porencephaly), suggesting that alternative approaches like pre-set ROIs might be needed for such cases.
- Future Directions:
- Prospective Studies: The authors advocate for future collaborative prospective studies to refine the model and further validate its impact on clinical practice.
- Comparison of HFA Measures: Further research is needed to determine whether HFA amplitude, coupling consistency, or both are more informative for SOZ localization and outcome prediction.
- Sharing Data: The authors express willingness to share their iEEG dataset with other researchers to test different HFA analytic methods.
- Wakefulness MI: The utility of MI during wakefulness for SOZ localization and outcome prediction remains to be investigated.
Summary updated