Automatic Detection of Scalp High-Frequency Oscillations Based on Deep Learning.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Scalp high-frequency oscillations (sHFOs) are crucial non-invasive biomarkers for epilepsy diagnosis.
• Current visual marking of sHFOs is laborious and subjective.
• Existing automated detectors struggle with artifact removal and clinical reliability.

Purpose of the Study:

• To develop a high-performance, automated detector for sHFOs using deep learning.
• To improve the accuracy and reliability of sHFO detection for clinical applications.
• To overcome limitations of existing single-dimensional analysis methods.

Main Methods:

• A deep learning approach combining 1D and 2D models was developed.
• An initial detection module identified candidate sHFOs.
• A weighted voting method integrated outputs from both models for final detection.

Main Results:

• The detector achieved high performance metrics: 83.44% precision, 83.60% recall, 96.61% specificity, and 83.42% F1-score.
• A kappa coefficient of 80.02% indicated strong agreement.
• The detector demonstrated stable performance across multi-center datasets, showing robustness and generalizability.

Conclusions:

• The proposed deep learning-based sHFO detector provides an accurate and robust method for epilepsy biomarker identification.
• Its high performance and generalizability suggest significant potential as a clinical decision support tool.
• This automated approach addresses the need for reliable and efficient sHFO analysis in epilepsy diagnosis.