| dc.description.abstract | Functional magnetic resonance imaging (fMRI)-based functional connectivity offers remarkable potential to non-invasively map individual differences in brain functional organisation to individual differences in behavioural phenotypes. Applied to neurodevelopmental and psychiatric conditions, this potential has renewed the search for neuroimaging-based biomarkers that have diagnostic and prognostic validity. Reproducible functional connectivity-based measures are key to achieving this goal. The overarching goal of this thesis is to improve the reproducibility of conventional and advanced functional connectomics-based measures to enhance their sensitivity to phenotypic differences of clinical relevance in neurodevelopmental and psychiatric conditions. I achieve this through four empirical studies using Open Science datasets as well as in-house data. All analysis code used in this thesis are publicly accessible via GitHub repositories: https://github.com/JRam02.
Using the functional connectome fingerprinting approach, Study 1 demonstrates that the test-retest fingerprint matching algorithm can serve as a simple and intuitive measure to compare, benchmark, and adjudicate between analysis pipelines. I show that this approach enables the identification of data pre- and post-processing parameters that optimise the detection of individual differences in the functional connectome. I also argue that fingerprint accuracy offers a number of advantages over the conventional approach to comparing pipelines ? the computation of test-retest reliability using the intraclass correlation coefficient (ICC). Across five datasets spanning childhood and adulthood derived from the Consortium for Replicability and Reproducibility (CoRR; N = 264, 7-43 years, mean age?SD = 19.4?5.6 years), I show that identification accuracies of up to 94% are achieved for pipelines that employ: (I) global signal regression; (II) finer-grained brain parcellations; (III) cortical regions, excluding subcortical and cerebellar structures; (IV) medial frontal and frontoparietal networks; (V) a focus on discriminative edges; and (VI) longer scan duration. Together, these factors form components of two pipelines that are optimised for reproducing the individual functional connectome. I suggest that a pipeline defined in this way should better enable the detection of robust and reproducible brain-behaviour relationships, particularly in developmental samples.
In Study 2, I address the trade-off between strict data quality requirements (particularly regarding head motion) and the large sample sizes required for robust and reproducible brain-behaviour associations. I propose the implementation of a machine learning technique referred to as ?bagging? or bootstrap aggregation at the timeseries level for estimating individual functional connectomes, enabling the salvage of ?high-motion? individuals that would otherwise be discarded in these analyses. Using resting-state data from CoRR and Healthy Brain Network (HBN; N = 423, 6-20 years, mean age?SD = 10.7?3.0 years), I show that deriving the functional connectome from bootstrap samples of low-motion timepoints can successfully differentiate an individual from a group and increases sample inclusivity. I further demonstrate that bagging yields robust univariate brain-behaviour associations (i.e., relationships between functional connectivity and age) even when as few as 60 ?clean? timepoints are retained (16% of the full timeseries; sampled with replacement over 500 iterations): the brain-behaviour associations obtained using bagging are comparable with those obtained using a standard (full timeseries) approach. The utility of bagging lies in its ability to rescue high-motion individuals by maximising participant inclusivity by ~4-30% of the high-motion individuals from CoRR and HBN; its application may reduce the tension between data quality vs. sample size requirements, mitigating resource wastage and the unethical practice of discarding large numbers of participants to ensure inclusive and reproducible brain-behaviour associations.
Study 3 builds on the work of Studies 1 and 2 by investigating whether the factors that maximise the detection of individual differences in the functional connectome, together with head motion strategies that maximise participant inclusivity and reduce motion artefacts also boost the predictive power of functional connectomes for behavioural phenotypes, using connectome-based predictive modelling (CPM). Using the HBN (N = 540, 6-21 years, mean age?SD = 10.7?3.0 years) and the components and strategies identified in Studies 1 and 2, I sought to strengthen connectome-based predictions and improve their generalisability across three domains: age, general psychopathology [CBCL], and intellectual ability [FSIQ]. I found that the data pre- and post-processing factors previously identified for improving the identifiability of individual functional connectomes ? (I) dimensionality of parcellation schemes; (II) focus on discriminatory functional networks such as the medial frontal and frontoparietal networks; and (III) head motion mitigation strategies including scrubbing and bagging ? neither improved nor disimproved the CPM-based predictions. Generalisability to an independent sample of non-Autistic individuals obtained from the Autism Brain Imaging Data Exchange (ABIDE; N = 439, 6-57 years, M = 16.2?6.7 years) was also generally unaffected, although finer parcellation schemes and bagging improved generalisability for FSIQ. Therefore, while the factors and strategies that I examined do not necessarily improve validity (i.e., strength of brain-behaviour relationships), the evidence suggests that they do strengthen reproducibility (i.e., generalisability of brain-behaviour relationships).
In Study 4, I take a different route to reproducibility: ecological validity. While resting-state fMRI conditions have been widely applied to study brain functions and dysfunctions across development, naturalistic paradigms may offer a more ecologically valid index of cognitive, emotional, and social processes. I propose a naturalistic analysis framework that employs a short emotionally evocative movie called ?The Present? (https://www.jacob-frey.com) to explore how functional brain responses are modulated by depressive symptomatology and emotion regulation strategies at multiple spatial and temporal scales in a sample of female adolescents experiencing subclinical depressive symptoms (N = 50, 13-20 years, mean age?SD = 17.0?1.9 years). In this study, I examined the (I) shared brain responses using intersubject correlation, (II) ?static? and dynamic functional connectivity patterns, and (III) signal complexity of the naturalistic fMRI signals using the Hurst Exponent (H), as well as their phenotypic relevance for depressive symptom severity and adaptive/maladaptive emotion regulation strategies. The findings suggest that depressive symptomatology and emotion regulation strategies shape the processing of affect-laden events in real-life scenarios.
Taken together, the findings of these four thesis projects highlight opportunities for improving the reproducibility of functional connectivity-based measures in developing populations, but also reveal several limitations that remain to be addressed. I discuss these findings in the context of several broader topics of concern within the neuroimaging field: sample size, the phenotype challenge, trade-offs of the resting-state fMRI approach, methodological variability, rate of developmental changes, and the quest for multimodal MRI approaches. I also consider several future avenues towards increased reproducibility, including meta-matching frameworks, multivariate techniques, multiverse pipelines, and sex differences. I conclude that a reproducibility lens remains central to the collective effort to discover functional connectivity-based biomarkers of clinical relevance, and ultimately, to improve outcomes for those with neurodevelopmental and psychiatric conditions. | en |
