Advancing EEG prediction with deep learning and uncertainty estimation.

Area of Science:

• Neuroscience and Artificial Intelligence
• Computational Neuroscience
• Machine Learning in Healthcare

Background:

• Deep learning (DL) offers potential in healthcare for disease detection and diagnosis.
• Lack of interpretability and complexity hinder DL adoption in critical healthcare predictions.
• Uncertainty estimation and explainability measures are crucial for building trust in DL systems.

Purpose of the Study:

• To investigate DL models for predicting sex from electroencephalography (EEG) data as a benchmark.
• To explore the use of DL ensembles for improved performance and interpretability.
• To evaluate the role of uncertainty estimation in enhancing DL model trustworthiness and performance.

Main Methods:

• Utilized InceptionNetwork and EEGNet models for sex prediction from EEG data.
• Employed DL ensembles, combining model variations, to enhance prediction accuracy.
• Implemented five-fold cross-validation for robust performance evaluation.
• Analyzed the relationship between frequency bands and sex prediction using a data-driven approach.

Main Results:

• A single InceptionNetwork model achieved 90.7% accuracy and an AUC of 0.947.
• The best DL ensemble reached 91.1% accuracy in sex prediction from EEG data.
• Uncertainty estimation via deep ensembles improved prediction performance.
• Models successfully classified sex across all frequency bands, revealing sex-specific features.

Conclusions:

• DL models, particularly ensembles with uncertainty estimation, show promise for EEG data analysis.
• Sex prediction from EEG serves as a valuable benchmark for developing interpretable and trustworthy AI in healthcare.
• Sex-specific neural features are present across all EEG frequency bands.