Predicting developmental age in young children by applying deep learning approaches to EEG data

Abstract

EEG-based age prediction models could be suitable indicators of brain maturational levels in children (from infancy till adolescence). In the past, traditional machine learning (ML) methods have been applied for predicting the age of young children using EEG data demonstrating the general feasibility of such an approach. Using an EEG data set (N=1,368,001 EEG epochs) from an extensive longitudinal study (children aged 11 to 47 months, N=304) we adapted state-of-the- art deep learning (DL) techniques to obtain improved age predictions for young children. This included implementing suitable tools to assess the uncertainty of the DL model, as well as tools to gain further insight into the models’ decision-making process (“explainable AI”). First, using feature extraction techniques and traditional ML tools, we created a reference age prediction model. We then implemented and compared seven DL regression architectures, trained on raw EEG data. To test the hypothesis that EEG-based age estimates reflect brain maturational level, we investigated if advanced/delayed EEG-based ages correlate with expressive and receptive vocabulary size. Finally, we investigated if the models’ age estimates are predictive of the presence or absence of recorded dyslexia predisposition. The best model, the cross-validated Encoder DL model, produced a mean absolute error of 4.82 months, a root mean squared error of 5.90 months, and an R-squared of 0.674. The brain age gaps (age estimate minus chronological age) were moderately to highly stable over time (0.534 ≤ r ≤ 0.835). Depending on the model, vocabulary sizes and brain age gaps have a weak to strong positive correlation at the age of 17, 23 (only expressive), and 35 months. A significant relationship between the brain age gap estimate and dyslexia predisposition was not found. We showed that DL models can outperform traditional ML models on infant age prediction using EEG data. DL models can extract features from raw EEG data, indicating the feature extraction and selection process can be avoided. For some DL models, the brain age gaps contain information about the subject’s development (vocabulary), depending on the model and age group. We presented an outlook on DL model explainability possibilities. Based on our findings, we expect that it is possible to use DL models to find other biomarkers in EEG data as well.