Evaluating Deep Learning Performance for P300 Neural Signal Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Signal Processing

Background:

• P300 event-related potential (ERP) signals are crucial neurological biomarkers for assessing cognitive functions, particularly in patients with neurological disorders.
• Accurate classification of P300 signals is essential, but existing models yield inconsistent results, lacking a consensus on optimal methods.
• Neurological disorder research requires reliable P300 signal analysis for cognitive function assessment.

Purpose of the Study:

• To evaluate and compare the performance of classic machine learning and novel deep learning methods for P300 signal classification.
• To assess classification performance in both within-subject and across-subject training paradigms.
• To determine the potential of deep learning models for generalized, subject-independent P300 classification in clinical settings.

Main Methods:

• Evaluated classic machine learning (Stepwise Linear Discriminant Analysis - SWLDA) and deep learning models (EEG-Inception).
• Utilized a dataset comprising P300 signals from 75 subjects.
• Compared model performance across within-subject and across-subject training and testing scenarios.

Main Results:

• Deep learning models achieved high F1 scores for attended event classification.
• In the within-subject paradigm, deep learning models did not outperform SWLDA.
• In the across-subject paradigm, EEG-Inception significantly outperformed SWLDA, demonstrating superior generalization capabilities.

Conclusions:

• Deep learning models show potential for P300 signal classification, particularly in across-subject scenarios, reducing the need for subject-specific calibration.
• SWLDA remains a strong performer in within-subject P300 classification tasks.
• The findings suggest deep learning models could offer a more generalized approach for clinical applications of P300 analysis.