Search

In recent decades, replication efforts in research have found that many findings are not reproducible. Many of these studies serve as the basis for others that might be relying on false assumptions. This replication crisis stands out in neurodevelopment research where heterogeneity in the typical human brain cannot, in most cases, be probed directly and relies on proxy measures of brain activity. This thesis develops three methodological frameworks for more robust research paradigms. I employ machine learning algorithms to navigate and optimise spaces of hidden variables, such as outcome variation between individual participants or data processing pipelines. The first framework builds a closed-loop experiment where an experimental space is explored automatically to maximise an individual's brain response. Generative modelling is used to create spaces of face stimuli to be explored in visual self-recognition. The framework is extended to EEG experiments with a mum-stranger paradigm run with infant participants. This allows the researcher to learn each individual's responses across many stimuli. The second framework builds a searchable space of different analysis. These spaces are used to model how robust each approach is within the multiverse of different analysis options. First, the multiverse of preprocessing pipelines is explored for functional connectivity data with the task of predicting brain age from adolescent developmental data. Second, a multiverse of predictive models is explored for an EEG face processing task predicting autism. The third framework is a normative modelling approach that uses state-of-the-art machine learning algorithms to model normal variability in brain structure. This approach generalises to different cohorts characterised by deviations from typical brain structure, detecting them as outliers. We illustrate its use by successfully predicting a neurodevelopmental psychiatric condition. This work intends to explore different avenues to build new gold standards in methodology that can improve the robustness of neurodevelopment and neuropsychiatry research.