Inter-site generalizability of EEG based age prediction algorithms in the preterm infant

Area of Science:

• Neuroscience
• Biomedical Engineering
• Developmental Pediatrics

Background:

• Brain age prediction using electroencephalogram (EEG) is a promising tool for assessing neurodevelopment in preterm infants.
• Site-specific variations in EEG data collection can significantly hinder the generalizability of predictive models.
• Developing robust methods to overcome these site differences is crucial for widespread clinical application.

Purpose of the Study:

• To develop and validate an EEG-based brain age prediction model for preterm infants that is robust to site differences.
• To investigate the impact of feature selection strategies on model performance across different data acquisition sites.
• To establish criteria for successful validation of brain age prediction models in multi-site settings.

Main Methods:

• A 'bag of features' approach combined with support vector regression (SVR) was employed for age prediction.
• Feature selection involved a filter-then-wrapper approach to identify reliable features across sites.
• Model training and validation were performed using EEG data from two distinct sites, with a focus on minimizing site-specific feature variations.

Main Results:

• An initial model trained on all features showed poor validation performance at the second site (MAE: 2.1 weeks vs. 1.0 week).
• Utilizing a restricted feature set, comprising features with similar distributions across sites, significantly improved validation accuracy (MAE: 1.1 weeks vs. 1.0 week).
• The restricted feature set approach resulted in a validated age predictor with no significant difference in prediction error between sites (p=0.68).

Conclusions:

• Site-dependent variations in EEG signals present a significant challenge for the generalizability of brain age prediction models.
• Employing a restricted feature set, carefully selected for cross-site consistency, is effective in overcoming these variations.
• Allowing minimal data leakage during feature selection can enhance the generalization of EEG interpretation methods for preterm infants, paving the way for universal application.