Multi-Center Validation Study of Automated Classification of Pathological Slowing in Adult Scalp...

Area of Science:

• Neuroscience
• Medical Technology
• Artificial Intelligence

Background:

• Pathological slowing in electroencephalogram (EEG) is crucial for diagnosing neurological disorders.
• Current visual inspection by experts is time-consuming and subjective.
• Automated detection systems are needed to improve EEG analysis efficiency and objectivity.

Purpose of the Study:

• To develop and evaluate automated systems for detecting pathological slowing in EEG.
• To compare the performance of Threshold-based Detection System (TDS), Shallow Learning-based Detection System (SLDS), and Deep Learning-based Detection System (DLDS).
• To assess system performance at channel-, segment-, and EEG-levels using rigorous cross-validation.

Main Methods:

• Development of three automated EEG slowing detection systems: TDS, SLDS, and DLDS.
• Evaluation across channel-, segment-, and EEG-levels using histogram-based aggregation of channel-level predictions.
• Validation using Leave-One-Subject-Out (LOSO) and Leave-One-Institution-Out (LOIO) cross-validation on diverse datasets.

Main Results:

• The Deep Learning-based Detection System (DLDS) demonstrated superior performance across all levels.
• DLDS achieved high balanced accuracy (BAC) in LOIO CV (up to 82.0%) and LOSO CV (up to 81.8%).
• Performance at channel- and segment-levels closely matched expert intra-rater agreement (IRA).

Conclusions:

• DLDS offers an efficient and accurate automated solution for pathological EEG slowing detection.
• The system can process 30-minute EEGs in just 4 seconds, significantly aiding clinical interpretation.
• Automated EEG analysis holds potential to assist clinicians in diagnosing neurological disorders more effectively.