The Kennard Principle, also known as the pediatric advantage, describes the seemingly better functional outcomes following brain injury in young children compared to adults. This principle is particularly prominent in the domain of language acquisition and development. Neuroimaging studies have demonstrated that language representation is bilateral and diffuse in early childhood, becoming increasingly left-lateralized and focal throughout adolescence. This suggests that extensive representation and/or topological redundancies may contribute to the pediatric advantage. Redundancy in neural circuits is crucial for flexible adaptation in both health and disease, potentially conferring network robustness to various perturbations, particularly during early development. This study utilizes a novel approach to investigate this topological redundancy. Brain networks are often represented as graphs, allowing for *in silico* attacks to simulate lesion impact and assess network robustness. Traditional approaches focus on the size of the largest connected component after node removal, defining the percolation point as the complete disintegration of the network. However, this approach has limitations. Therefore, this study utilizes the size of the second largest connected component as an alternative measure of network resilience and employs functional data analysis (FDA) to overcome issues related to thresholding in brain network analyses.

Previous research supports the concept of the pediatric advantage in language following brain injury. Studies have shown that unilateral lesions cause different effects on language production in children compared to adults, with children demonstrating a remarkable capacity for recovery. Neuroimaging studies have highlighted the bilateral and diffuse nature of the language network in young children, gradually shifting towards left lateralization and focality with age. The existing literature suggests that the increased resilience in children might be attributed to extensive representation or topological redundancies within their neural circuits. These redundancies act as a buffer, enabling flexible adaptations to both neurological and environmental challenges, particularly during early development. Network analysis, which represents brain connectivity as a graph, provides a robust method for investigating these topological properties, with *in silico* lesion studies offering a means to assess resilience against targeted or random damage. However, existing methods for assessing network resilience using such simulated lesions have limitations in finite systems like the brain, particularly regarding the selection of a critical threshold for analyzing the connectivity data.

This study employed magnetoencephalography (MEG) data from 85 typically developing children and adolescents aged 4 to less than 19 years. High-quality MEG and MRI data were obtained from 82 participants (45 female). All participants were native English speakers with no history of neurological or hearing impairment, speech or language deficits, or learning disabilities. The study used a stories listening paradigm, where participants listened to child-friendly stories interspersed with speech-shaped noise. MEG data were acquired using a 275-channel whole-head MEG system at 1200 Hz. Structural MRI was acquired after MEG, allowing for precise coregistration. MEG data were preprocessed using FieldTrip in MATLAB, including bandpass filtering, noise suppression, and independent component analysis (ICA) to remove artifacts. Realistic single-shell source models were constructed from segmented T1-weighted images, and source activity was estimated using a linearly constrained minimum variance beamformer. Functional connectivity was assessed using weighted phase lag index (wPLI) for narrow frequency bins, aggregated to estimate total coupling. Whole-brain and language networks (defined using NeuroSynth meta-analysis) were created. In silico attacks involved removing nodes randomly or based on eigenvector centrality (EC) and betweenness centrality (BC) at various network densities. The percolation point was determined by the peak size of the second largest connected component. Functional data analysis (FDA) was used to model the relationship between age, sex, handedness, mean node distance (a covariate to account for brain size differences), and percolation point across densities. The penalized flexible functional regression (PFFR) approach was employed. Significance testing used the likelihood ratio test, bootstrap resampling, and a conservative p-value threshold of 0.001.

The analysis revealed significant effects of age on percolation point for random and betweenness centrality-based attacks, but not for eigenvector centrality-based attacks. For both random and betweenness centrality attacks on the whole-brain network, younger children exhibited higher percolation points, indicating greater network resilience (Figures 1 and 2). This age effect was consistent across a range of network densities (below 15% for random attacks; across all densities for betweenness centrality attacks). Similar findings were observed for the stories network (Figures 4 and 5), with younger participants showing more resilience to both random and betweenness centrality attacks. Visualization of removed nodes at 5% density showed a pattern where younger children had more consistently removed regions across participants (group-level hubs), demonstrating a more distributed pattern of critical hubs, particularly in posterior regions (Figure 3). Older participants exhibited fewer consistent removal regions, with critical hubs becoming more focal (Figure 3, 6). This suggests that young children's language networks rely on a more distributed pattern of connectivity, while older children's networks shift towards greater focality.

This study demonstrates that network vulnerability increases with age in the language network, supporting the concept of the pediatric advantage. The age effect is more prominent for betweenness centrality-based attacks, indicating an increasing importance of domain-specific hubs with age, while domain-general connector hubs remain relatively stable across development. This finding is consistent with the developmental changes in brain connectivity, characterized by the pruning of short-range connections and increased long-range connections. The increased efficiency in information transfer resulting from these developmental changes appears to come at the cost of increased network vulnerability. The observed shift from distributed critical hubs in younger children to more focal critical hubs in older participants may reflect the increasing reliance on intrahemispheric reorganization in older children. These findings have implications for understanding the neurobiological underpinnings of the pediatric advantage and for guiding surgical decisions in children, potentially shifting from a focus on functional localization to a network-based approach that considers tissue redundancy and potential for functional compensation.

This study provides strong evidence of increased vulnerability of the language network from childhood to adolescence. The findings highlight the importance of considering both domain-specific and domain-general hubs in assessing network resilience and the capacity for functional compensation after injury. Future research should focus on single-subject level analyses of critical hubs to refine our understanding of network plasticity and inform surgical planning in children.

The study utilized a relatively homogeneous sample of typically developing children and adolescents. Future studies should explore the resilience of brain networks in diverse populations and across different cognitive domains. The *in silico* lesion method provides a simplified model of brain injury; real-world lesions are more complex. Finally, while FDA helps mitigate thresholding issues, the choice of specific analysis parameters (spline order, number of basis functions) might influence the results.