Classification of the Epileptic Seizure Onset Zone Based on Partial Annotation

Area of Science:

• Neuroscience
• Medical Informatics
• Signal Processing

Background:

• Epilepsy diagnosis relies on expert visual analysis of long-term intracranial electroencephalogram (iEEG) data to identify the seizure onset zone (SOZ).
• This manual process is time-consuming, challenging, and requires extensive experience, highlighting the need for automated diagnostic aids.
• Machine learning (ML) offers potential solutions for improving the efficiency and accuracy of SOZ identification.

Purpose of the Study:

• To develop and evaluate an ML-based method for accurate and efficient classification of seizure onset zone (SOZ) and non-SOZ iEEG data.
• To compare the effectiveness of various feature extraction techniques, including traditional methods and advanced transforms, for SOZ detection.
• To investigate the utility of positive unlabeled (PU) learning to minimize the annotation burden on clinical experts.

Main Methods:

• Intracranial electroencephalogram (iEEG) data were segmented into 20-second intervals.
• Feature extraction was performed using filtering, entropy, short-time Fourier transform (STFT), wavelet transform (WT), and empirical mode decomposition (EMD).
• Classification models including support vector machine (SVM), fully connected neural network (FCNN), and convolutional neural network (CNN) were trained and evaluated, incorporating positive unlabeled (PU) learning.

Main Results:

• The proposed method achieved high-performance classification of SOZ and non-SOZ data using ML models.
• Feature extraction methods were compared to identify the most discriminating features for SOZ detection.
• The integration of PU learning significantly reduced the expert annotation workload while maintaining high classification accuracy.

Conclusions:

• The developed ML approach provides a high-accuracy diagnostic aid for identifying the seizure onset zone (SOZ) from iEEG data.
• Feature engineering and advanced ML models, combined with PU learning, offer a promising strategy to reduce expert workload in epilepsy diagnosis.
• This method has the potential to improve the efficiency and consistency of SOZ identification in clinical practice.