Children who suffer from brain insults (i.e., traumatic brain injury (TBI), chemotherapy and radiation treatment for brain tumors) are susceptible to late-emerging cognitive sequelae. Even with similar neurological risk variables, variability in long-term cognitive outcomes remains an area of investigation for researchers of acquired brain injury. Given the potential for genetic factors to influence response to chemoradiation, researchers have examined associations between germline, inherited, single nucleotide polymorphisms (SNPs), and neurocognitive outcomes for cancer survivors. Children who sustain an uncomplicated mild TBI generally recover without long-term neuropsychological consequences. However, TBI survivors have overlapping mechanism categories with cancer survivors through secondary injury variables that can be influenced by genomic variation (e.g., oxidative stress and neuroinflammation). Furthermore, the study of genomic vulnerability is limited in heterogenous groups of pediatric TBI survivors. This study aims to identify associations between genotype and long-term neurocognitive outcomes for acquired brain injury survivors by utilizing machine learning to uncover pathophysiological similarities and differences between groups.

Fourteen brain tumor survivors, 139 traumatic brain injury survivors, and 63 healthy, age-matched controls completed the Letter N-back task to obtain performances on core neurocognitive skills (attention, working memory, and processing speed). Ten targeted genotypes were examined across five pathophysiological pathways (neurotransmission, oxidative stress, neuroinflammation, plasticity, growth and repair, and folate metabolism). Data were trained and tested utilizing three regression machine learning models. Mean estimated error and R2 were generated for each neurocognitive outcome. A feature importance score for models with positive variance was generated to determine how predictive a given SNP is for neurocognitive outcomes.

Genotype only accounted for a small amount of variance in cognitive outcomes when all clinical groups were combined. The mean absolute error for the best-fitting models from analyses where all groups were combined decreased when groups were examined separately; however, the differences in model R2 values were not significant. The relationship between brain tumor survivors and processing speed performance depended on genotype. Two SNPs had positive feature importance at the interaction level (rs58225473 and rs1801394). These SNPs are located on the CACNB2 and MTR genes and have functional consequences for neurotransmission and folate metabolism. Models of traumatic brain injury survivors did not explain positive variance and could not be examined for feature importance. Additionally, even when removing the only mechanism of action that should not be relevant for TBI survivors (folate metabolism polymorphisms), the TBI models still did not explain positive variance.

Findings of the importance of two key SNPs on MTR and CACNB2 genes align with recent systematic reviews, which found associations between these polymorphisms and neuropsychological outcomes in more than one group or cohort of pediatric cancer survivors. Models for TBI survivors were limited by the heterogeneity of the group and ceiling effects on performance. An understanding of genetic vulnerabilities influenced by treatment and injury-related factors in acquired brain injury will inform our understanding of the developing and recovering childhood brain. The current study is an initial contribution to this goal and highlights the utility of machine learning methodology for future studies that examine the influence of genetic heterogeneity in pediatric acquired brain injury.