The Role of EEG Microstates in Predicting the Therapeutic Outcomes of Oxcarbazepine in Newly Diagnosed Focal Epilepsy Patients

Introduction

Background

Focal epilepsy is the most common type of epilepsy, accounting for about 60% of all epilepsy cases. The selection of antiepileptic drugs (AEDs) varies depending on the type of epilepsy. In the treatment of focal epilepsy, oxcarbazepine (OXC) is widely used. However, while oxcarbazepine can achieve seizure freedom in about 65% of patients, a significant portion still fails to achieve satisfactory therapeutic outcomes. Electrophysiological monitoring techniques, such as electroencephalography (EEG), play a crucial role in the diagnosis and management of epilepsy.

Research Purpose

A microstate is an EEG pattern reflecting the spatiotemporal characteristics of brain electrical activity. Previous studies have shown that AEDs can affect EEG signals of the brain, but research on oxcarbazepine remains limited. Moreover, studies indicate that transient states could become powerful biomarkers for epilepsy. Therefore, this study aims to investigate the EEG microstates of newly diagnosed focal epilepsy patients before oxcarbazepine treatment and to predict the therapeutic outcomes using extracted microstate features.

Research Team and Publication Information

This study was completed jointly by Rong Rong, Zhang Run-kai, Xu Yun, Wang Xiao-yun, Wang Hai-xia, and Wang Xiao-shan. Rong Rong and Xu Yun are from Nanjing Drum Tower Hospital, Zhang Run-kai and Wang Hai-xia are from Southeast University, and Wang Xiao-shan is from Nanjing Brain Hospital. The paper was published on May 23, 2024, in the “Seizure: European Journal of Epilepsy.”

Research Methods

Subjects and Data Collection

The subjects included 25 newly diagnosed focal epilepsy patients (13 females), aged between 12 and 68 years. Based on the first follow-up results, the subjects were divided into seizure-free (SF) and non-seizure-free (NSF) groups. All patients underwent long-term scalp EEG recordings before starting monotherapy with oxcarbazepine. The inclusion criteria were: patients had not used antiepileptic drugs at initial diagnosis, were treated exclusively with oxcarbazepine for at least six months, and had available follow-up records.

Research Procedure

1. Data Preprocessing:

   • All patients underwent 19-channel EEG recordings (sampling rate of 256Hz or 512Hz) before starting oxcarbazepine treatment.
   • Preprocessing was performed using MNE-Python software, including filtering, re-referencing, and artifact removal.

2. Microstate Analysis:

   • Using a modified K-means clustering method, the topographic maps of scalp potentials around the global field power peaks of the EEG were clustered into four representative microstates.
   • Temporal and spatial parameters of each microstate were extracted, including duration, coverage, occurrence frequency, and transition probability.

3. Machine Learning Prediction:

   • Extracted microstate features were used to build machine learning models, including Logistic Regression (LR), Naive Bayes (NB), and Support Vector Machine (SVM).
   • Four-fold cross-validation was used to evaluate model performance, and the ADASY algorithm was employed for data augmentation to balance the datasets.

Data Analysis

• Independent sample T-tests and Wilcoxon rank-sum tests were used to compare microstate parameters between different groups.
• Fisher’s exact test was used to assess the associations between binary variables, such as the presence of temporal lobe epilepsy (TLE) and prognosis.
• The predictive performance of machine learning models was evaluated using Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC).

Experimental Results

Main Results

1. Compared to the SF group, the NSF group showed significantly higher durations, occurrence frequencies, and coverages for Microstate 1 (MS1).
2. The global explained variance and temporal correlation of Microstate 2 (MS2) were significantly higher in the NSF group than in the SF group.
3. The SF group showed higher transition probabilities from MS2 to MS4, while the NSF group exhibited higher transition probabilities from MS2 and MS3 to MS1.

Predictive Results

When using microstate features for prediction: - The LR model achieved an average AUC of 0.95 on Feature Set I. - The SVM model achieved an average AUC of 0.84 on Feature Set III. - The NB model achieved an average AUC of 0.70 on Feature Set III.

Discussion

This study is the first to use EEG microstate analysis to examine the brain activity of newly diagnosed focal epilepsy patients before oxcarbazepine treatment and to show that microstate features can significantly predict therapeutic outcomes. The results indicate that high activity levels of MS1 in the NSF group can serve as a marker for poor outcomes, and characteristics of MS2 and MS4 also show important predictive value.

Clinical Implications

EEG microstate analysis can serve as a tool to predict the therapeutic outcomes of oxcarbazepine, helping doctors choose more suitable treatment plans, improving therapeutic efficacy, and reducing unnecessary treatment costs and psychological burden. This method’s advantage lies in capturing the spatial-temporal characteristics of brain activity, providing important neurophysiological information.

Conclusion

This study demonstrated the potential of EEG microstates in predicting the therapeutic outcomes of oxcarbazepine in newly diagnosed focal epilepsy patients. By extracting microstate features and using machine learning models, the research team successfully achieved high-accuracy predictions of therapeutic outcomes. Despite the limited data, the results underscore the importance of EEG microstates as biomarkers for antiepileptic drug effects and provide a foundation for future machine learning research.

Research Significance and Future Directions

• Incorporating small samples into larger-scale studies or external validations to confirm the research results.
• Exploring high-density EEG or other techniques such as magnetoencephalography to provide new perspectives on microstates.
• Further combining multiple bioelectrical and imaging features to enhance the robustness and predictive ability of machine learning models.