Neural feature spaces: characterizing the geometry of brain activity across development

Abstract

The human brain undergoes profound changes from childhood to adulthood. These changes are foundational to cognitive, affective and social development, and measuring them is essential for identifying atypical development. A common approach consists of analyzing metrics derived from brain activity, such as spectral power, aperiodic activity and signal complexity. However, individual metrics are sensitive to external factors unrelated to development, contributing to inconsistent findings in the literature. Moreover, these metrics are not independent; they show some degree of intrinsic redundancy, which itself evolves across development. Here, we propose a novel approach that focuses on the relationships between neural features rather than their values, defining a "neural feature space" that captures the geometry of brain activity. We analyzed the neural feature spaces of children (4-12 years) and adults (30-45 years), based on 128-channel EEG recordings during naturalistic movie viewing. Specifically, we computed a range of spectral, aperiodic and complexity features and quantified pairwise distances between them using latent variable modeling. Neural feature spaces were significantly different between age groups. Notably, high frequency bands were less differentiated in children and distances between spectral and complexity features differed. Crucially, distances were highly stable across tasks, in contrast to feature values, which varied substantially. These findings suggest that the geometry of neural feature spaces provides robust, interpretable markers of neurodevelopment, offering a complementary approach to feature-based analyses.