An automatic patient-specific seizure onset detection method in intracranial EEG based on incremental nonlinear...

Area of Science:

• * Neuroscience
• * Signal Processing
• * Machine Learning

Background:

• * Epileptic seizure detection relies on analyzing electroencephalographic (EEG) signal morphology and spatial distribution.
• * Accurate and automated seizure onset detection is crucial for patient management and research.
• * Existing methods may require extensive training data and human intervention.

Purpose of the Study:

• * To develop a novel incremental learning scheme for automatic patient-specific seizure onset detection.
• * To utilize nonlinear dimensionality reduction (NDR) for efficient feature extraction from EEG signals.
• * To enable accurate seizure onset identification in long-term intracranial EEG (iEEG) recordings.

Main Methods:

• * Employed Local Tangent Space Alignment (LTSA), a nonlinear dimensionality reduction technique, to process features extracted via Continuous Wavelet Transform (CWT).
• * Developed an unsupervised incremental learning scheme to update a one-dimensional manifold representing seizure onset dynamics.
• * Trained the initial manifold using iEEG recordings with a single seizure onset, allowing sequential updates with unlabeled data.

Main Results:

• * Achieved high performance on iEEG data from 21 patients (193.8h, 82 seizures).
• * Demonstrated average sensitivity of 98.8%, low false positive rates (0.24/h uninteresting, 0.25/h interesting), and a detection delay of 10.8s.
• * Showcased the method's effectiveness in off-line seizure detection with minimal human intervention.

Conclusions:

• * The proposed incremental learning scheme offers a simple and accurate approach for patient-specific seizure onset detection.
• * The method requires minimal training data and has potential for real-time applications.
• * The unsupervised incremental learning framework can identify novel EEG patterns, including different seizure onset types.