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

Abstract

Objective: To assess and overcome the effects of site differences in EEG -based brain age prediction in preterm infants. 
Approach: We used a ‘bag of features’ with a combination function estimated using support vector regression (SVR) and feature selection (filter then wrapper) to predict post-menstrual age (PMA). The SVR was trained on a dataset containing 138 EEG recordings from 37 preterm infants (site 1). A separate set of 36 EEG recordings from 36 preterm infants was used to validate the age predictor (site 2). The feature distributions were compared between sites, and training used only features that were not significantly different between sites. The mean absolute error between predicted age and PMA was used to define the accuracy of prediction. Successful validation was defined as no significant differences in error between site 1 (cross-validation) and site 2.
Main results: The age predictor based on all features and trained on site 1 was not validated on site 2 (p < 0.001; MAE site 1 = 1.0 weeks, n = 59 vs MAE site 2 = 2.1 weeks, n = 36). The MAE was improved by training on a restricted features set (MAE site 1 = 1.0 weeks, n = 59 vs MAE site 2 = 1.1 weeks, n = 36), resulting in a validated age predictor (p = 0.68). The selected features closely aligned with features selected when trained on a combination of data from site 1 and site 2.
Significance: The ability of EEG classifiers, such as brain age prediction, to maintain accuracy on data collected at other sites may be challenged by unexpected, site-dependent differences in EEG signals. Permitting a small amount of data leakage between sites improves generalization, leading towards universal methods of EEG interpretation in preterm infants.

Emergence of a synergistic scaffold in the brains of human infants

“…The combination was trained within a 10-fold cross validation where approximately 90% of cortical signals were included in a training set and 10% of cortical signals were left out for testing; a process that was repeated 10 times until all data had been tested on (due to the presence of multiple EEG recordings and twins in the dataset, cross-validation selections were based on mother's ID number). Within each training fold, feature selection was applied using a hybrid filterwrapper approach to reduce the dimensionality of the input feature vector [101]. As a first stage, only features with a significant correlation (corrected for multiple comparisons; Benjamini-Hochberg procedure) with age were selected (filter stage).…”