Katherine Birch

Human brain structure datasets are regularly small and lack a representative sample of target phenotypes. Augmenting these is often challenging due to highly complex patterns of connectivity, and this issue is particularly emphasized when there are significant target class imbalances. We introduce a model that enables the generation of novel data instances and data exploration. Specifically, we consider the case of preterm birth, where datasets include very few preterm individuals. We present a diffusion-style flow-matching framework, whereby conditioning on continuous gestational age (GA), the model learns the underlying geometry of the brain and can reproduce differences in connectivities for infants born at varying numbers of weeks. This approach is inspired by the brain's fundamental capacity for self-organization. Moreover, to understand the real implications of varying GA on the organization of the developing brain, we integrate a dynamic hypergraph layer. This allows the model to dynamically learn the higher-order dependencies that evolve in the brain, facilitating the generation of biologically plausible structural topologies. To ensure appropriateness and sufficiency in the generated networks, we utilize biologically informed losses. Additionally, we compare the generation with key graph metrics commonly employed in neurobiological studies of brain structure. Our model results demonstrate biologically realistic connectivity patterns discovered through the dynamic hypergraph approach. The code is publicly available at: https://github.com/Katherine-Birch/Generative-Flow-Matching-Modeling-of-the-Neurodevelopmental-Connectome-via-Dynamic-Hypergraphs.This is the author's version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record will be published in the Proceedings of the 32nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '26), doi: https://doi.org/10.1145/3770855.3819022

Preterm births have been associated with altered neurological development for neonatal infants, this has been implicated in certain neurodevelopmental conditions in later life. Advances in brain imaging methods, such as Magnetic Resonance Imaging, have allowed for the analysis of physical connectivity of brain matter in infants shortly after birth. However, commonly used methods of investigating such data rely on a brain network analysis, traditionally based on graph-theoretical approaches, which may fail to capture complex patterns involving both local and global network structures and spatial information. Furthermore, many previous studies of infant brain data rely on a priori selection of specific graph connectivity measures. We propose employing machine learning models such as logistic regression and Graph Neural Networks to provide a data-driven approach for classifying preterm and term brain networks at birth. We utilize fuzzy logic, and explainability methods including Shapley Additive Explanations to identify influential regions and connections in decision making. In our analysis, brain regions are represented as spatially embedded nodes, with edges representing strength of structural connections between areas. Using this setup, our model achieves a binary classification accuracy of 88.57%. This performance is further enhanced using a fuzzy boundary between preterm and term classes, achieving an accuracy of 96.19%. This demonstrates that the model can be assisted particularly by adding context to 'near-term' born infant cases. These analyses highlight important connections, and key nodes including deep brain structures which are broadly consistent with biological literature.