Multivariate resting-state EEG markers differentiate people with epilepsy and functional seizures

This study demonstrates that multivariate resting-state EEG network measures, particularly when enhanced by epoch-wise averaging, can significantly distinguish between non-lesional epilepsy and functional dissociative seizures with a maximum balanced accuracy of 67.5%, offering a potential clinical tool for improving diagnostic accuracy prior to treatment initiation.

The Gist

The Big Problem: The "Look-Alike" Seizures

Imagine you have two very different types of storms. One is a real thunderstorm (Epilepsy), caused by a sudden electrical short-circuit in the brain. The other is a dramatic play-acting storm (Functional/Dissociative Seizures, or FDS), where the body acts like it's having a storm, but there is no electrical short-circuit happening.

To the naked eye, they look exactly the same. Patients shake, lose consciousness, and need help. But the treatments are completely different. If you treat a "play-acting" storm with anti-seizure drugs, it doesn't work and can be harmful. If you treat a "real" storm with therapy alone, the patient remains at risk.

Doctors often struggle to tell them apart. Sometimes, the standard "weather report" (the EEG test) looks perfectly clear for both, leaving the doctor guessing. This paper asks: Can we use a smarter, more complex way of reading the brain's "weather" to tell these two storms apart?

The New Tool: Listening to the "Conversation"

Usually, when doctors look at an EEG, they listen to individual microphones (electrodes) on the scalp to see if one is screaming (spiking). This study, however, decided to stop listening to the microphones individually and started listening to how the microphones talk to each other.

Think of the brain not as a collection of isolated radio stations, but as a giant, busy office.

• Old Method (Univariate): Checking if one specific employee is shouting.
• New Method (Multivariate/Network): Checking the flow of conversation between all the employees. Are they all whispering to each other? Is the boss talking to the intern too much? Is the whole office moving in a synchronized, chaotic dance?

The researchers looked at the "resting state"—when the patients were just sitting quietly with their eyes closed. Even though the EEG looked "normal" to the human eye, the computer looked for subtle patterns in how different parts of the brain were connected.

The Experiment: Training a Digital Detective

The researchers gathered data from 148 people who were suspected of having seizures but hadn't been diagnosed yet.

• Group A: 75 people who actually had Epilepsy.
• Group B: 73 people who had Functional Seizures.

They fed this data into a computer (Machine Learning) and asked it to act as a Digital Detective. They tried many different "detective styles" (algorithms) to see which one could best spot the difference between the two groups.

The Key Trick:
They found that if they took a "snapshot" of the brain's conversation for a few seconds, it was a bit noisy and unreliable. But, if they took many snapshots and averaged them out (like taking a long-exposure photo to smooth out the blur), the detective got much better at solving the case.

The Results: The Detective Gets It Right (Mostly)

The computer detective was able to distinguish between the two groups significantly better than random guessing.

• The Score: The best detective got it right about 67.5% of the time. (Random guessing would be 50%).
• The Bias: The detective was very good at spotting the Real Storm (Epilepsy) (getting it right 82% of the time). However, it was a bit worse at spotting the Play-Acting Storm (FDS) (getting it right only 53% of the time).

What does this mean?
If the computer says, "This person has Epilepsy," you can be fairly confident. But if it says, "This person has FDS," you shouldn't be 100% sure yet. The tool is better at confirming the presence of the electrical storm than confirming the absence of it.

Why This Matters

1. No More Guessing Games: Currently, if a patient has a seizure and the EEG looks normal, doctors often have to wait weeks or months to see if the patient has a seizure during a video recording in a hospital. This tool could give a "second opinion" much faster, right when the patient first walks in.
2. Safety: It helps prevent giving the wrong medication to people who don't need it.
3. The "Network" Insight: It proves that even when the brain looks "quiet" and normal, the way the brain parts connect to each other holds a secret signature that tells us what's really going on.

The Catch (Limitations)

• It's not perfect yet: It's not a magic wand that gives a 100% diagnosis. It's a "decision support tool" to help doctors, not replace them.
• It's better at finding Epilepsy: As mentioned, it's great at saying "Yes, this is Epilepsy," but less good at saying "No, this is definitely FDS."
• More testing needed: This was a study on existing data. Before this can be used in a real hospital, it needs to be tested on new, fresh patients to make sure it works everywhere.

The Bottom Line

This paper is like finding a new pair of X-ray glasses for brain waves. Even when the brain looks normal on a standard check-up, these glasses can see the subtle "traffic patterns" of the brain's electrical signals. They can't solve every mystery yet, but they are a huge step forward in helping doctors tell the difference between a real electrical storm and a dramatic performance, ensuring patients get the right help sooner.

Technical Summary

1. Problem Statement

Distinguishing between non-lesional epilepsy and functional/dissociative seizures (FDS) is a significant clinical challenge. Misdiagnosis leads to delayed treatment, inappropriate medication, and increased morbidity.

• Current Limitations: Routine clinical EEGs often appear visually normal in both conditions. Previous attempts to use quantitative EEG (qEEG) relied on univariate features (single-channel metrics), which showed limited diagnostic validity when controlling for confounding factors (e.g., medication, comorbidities).
• Hypothesis: Diagnostic information may be hidden in multivariate features, specifically functional network measures (connectivity and topology), which capture the complex dynamics of brain synchrony better than single-channel metrics.

2. Methodology

Dataset and Participants

• Sample: 148 adults (75 epilepsy, 73 FDS) with suspected seizure disorders.
• Inclusion Criteria: Age- and sex-matched; non-lesional (normal MRI/CT); medication-free at time of EEG; no major neurological/psychiatric comorbidities.
• Data: Eyes-closed, resting-state EEG (21 channels). Segments were visually inspected to ensure they were "normal" (no epileptiform abnormalities or major artifacts).
• Subgroup: A subset of 102 participants (57 epilepsy, 45 FDS) was used for specific analyses requiring a minimum of four EEG epochs per individual to ensure feature stability.

Signal Processing & Network Construction

1. Preprocessing: Bandpass filtering (0.5–70 Hz), artifact removal (ICA), and re-referencing.
2. Frequency Band: Analysis focused on the 6–9 Hz range (referred to as "low alpha," overlapping classical theta).
3. Network Metrics:
   • Phase-Locking Factor (PLF): Calculated pairwise between electrodes to measure phase consistency.
   • Lag: Measured directionality of phase differences.
   • Surrogate Data: 99 surrogate epochs generated via iAAFT to establish a null distribution.
   • Edge Definition: Edges were retained only if PLF exceeded the 95th percentile of surrogates. Indirect paths were pruned to ensure direct connectivity.
   • Weighting: Edges were weighted and normalized to set mean node strength to 1.

Feature Extraction

Six functional connectivity-based network measures were calculated:

1. Mean Clustering Coefficient: Local interconnectivity.
2. Mean Functional Connectivity (MFC): Global coupling strength (raw PLF average).
3. Network Efficiency: Global information transfer efficiency.
4. Trophic Incoherence: Measure of network hierarchy/directionality.
5. Mean Betweenness Centrality: Node importance in shortest paths.
6. Critical Coupling: Theoretical threshold for seizure emergence based on Kuramoto oscillator models.

Machine Learning Framework

• Feature Selection: Nested backward elimination using a Naïve Bayes classifier identified a 3-feature subset as optimal: Network Efficiency, Trophic Incoherence, and Mean Functional Connectivity.
• Validation Strategy: Nested Cross-Validation (10-fold outer, 10-fold inner) to prevent data leakage and optimize hyperparameters.
• Model Configurations: Tested 14 model types (e.g., SVM, Random Forest, XGBoost, k-NN) with variations in:
  • Dimensionality reduction (PCA, Spectral Embedding).
  • Data transformation (Box-Cox).
  • Epoch averaging (1 vs. ≥4 epochs).
• Metric: Balanced Accuracy (average of sensitivity for Epilepsy and sensitivity for FDS) to handle class balance.

3. Key Results

Classification Performance

• Best Configuration: SVM with Radial Basis Function (RBF) kernel, using the 3-feature subset, no dimensionality reduction, and averaging over ≥4 epochs.
• Balanced Accuracy: 67.5% (significantly above chance, which was ~52% via permutation testing).
• Sensitivity Breakdown:
  • Epilepsy Sensitivity: 81.8% (High).
  • FDS Sensitivity: 53.3% (Low/Moderate).
• Consensus: The top three models (SVM-RBF, Random Forest, k-NN) agreed on the diagnostic label for 77.5% of participants.

Feature Stability & Model Choice

• Epoch Averaging: Using ≥4 epochs improved accuracy from ~62.6% to 67.5%, suggesting that temporal stability of network features is crucial.
• Dimensionality Reduction: Did not improve performance; the raw 3-feature space was sufficient.
• Model Type: Non-linear models (SVM-RBF, Random Forest, k-NN) significantly outperformed linear models (Logistic Regression, LDA), indicating complex, non-linear relationships between network topology and diagnosis.

Subgroup Analysis

• Robustness: Performance remained high for patients with completely normal EEGs (70.7%) and those with epilepsy-specific abnormalities elsewhere in the recording (68.1%).
• Limitations: Performance dropped for patients with non-specific abnormalities (e.g., slowing) elsewhere in the recording (61.5%) and for those with comorbidities (59.3%), driven by reduced FDS sensitivity.
• Frequency Bands: While 6–9 Hz was optimal, Delta and Theta bands also showed promise (48–60% accuracy), whereas Alpha and Beta performed lower.

4. Key Contributions

1. First Direct Differentiation: This is the first study to directly discriminate between non-lesional epilepsy and FDS using multivariate network measures from visually normal, medication-free resting-state EEG.
2. Diagnostic Validity: Establishes that network markers have clinical validity for estimating post-test probability, particularly for ruling in epilepsy (high sensitivity).
3. Methodological Rigor: Demonstrates the necessity of epoch-wise averaging to stabilize features and the superiority of non-linear classifiers for this specific diagnostic problem.
4. Feature Selection: Identifies a minimal, robust set of three network metrics (Efficiency, Trophic Incoherence, MFC) that outperform the full set of six.

5. Significance and Limitations

Clinical Significance

• Decision Support: These markers can serve as a decision-support tool to aid non-specialists in acknowledging diagnostic uncertainty and optimizing referrals to tertiary centers.
• Specificity: The markers are highly specific to epilepsy (high sensitivity for epilepsy, low for FDS). The authors explicitly state these should not be interpreted as a positive marker for FDS, as FDS sensitivity remains near chance levels.
• Timing: The method works prior to treatment initiation, addressing a critical window where misdiagnosis is most damaging.

Limitations & Future Directions

• FDS Sensitivity: The low sensitivity for FDS (53.3%) limits the tool's ability to confirm FDS. The authors suggest FDS may lack distinct interictal network signatures or that current univariate/multivariate approaches are insufficient for this group.
• Comorbidities: Performance degrades in patients with comorbidities, suggesting the need for models that account for psychiatric/neurological overlap.
• Retrospective Design: The study is retrospective; future work requires independent, prospective validation in multi-center trials to confirm generalizability.
• Sample Size: While well-matched, the sample size (N=148) is moderate, and the best results rely on a subset (N=102).

Conclusion: The study provides strong evidence that multivariate network analysis of resting-state EEG can differentiate epilepsy from FDS with accuracy significantly above chance, offering a promising avenue for objective diagnostic support, particularly for identifying epilepsy in visually normal EEGs.