Mental health in the UK Biobank: A roadmap to self‐report measures and neuroimaging correlates

Abstract

The UK Biobank (UKB) is a highly promising dataset for brain biomarker research into population mental health due to its unprecedented sample size and extensive phenotypic, imaging, and biological measurements. In this study, we aimed to provide a shared foundation for UKB neuroimaging research into mental health with a focus on anxiety and depression. We compared UKB self‐report measures and revealed important timing effects between scan acquisition and separate online acquisition of some mental health measures. To overcome these timing effects, we introduced and validated the Recent Depressive Symptoms (RDS‐4) score which we recommend for state‐dependent and longitudinal research in the UKB. We furthermore tested univariate and multivariate associations between brain imaging‐derived phenotypes (IDPs) and mental health. Our results showed a significant multivariate relationship between IDPs and mental health, which was replicable. Conversely, effect sizes for individual IDPs were small. Test–retest reliability of IDPs was stronger for measures of brain structure than for measures of brain function. Taken together, these results provide benchmarks and guidelines for future UKB research into brain biomarkers of mental health.

Population modeling with machine learning can enhance measures of mental health - Open-Data Replication

“…Bijsterbosch et al, 2020;. For example, network amplitudes have higher test-retest reliability than connectivity information (Dutt et al, 2021), and have been shown to capture individual differences in behavioral traits (J. Bijsterbosch et al, 2017;Miller et al, 2016).…”

“…Importantly, the phenotype of age is known without error (apart from potential database entry mistakes), whereas neuroticism and fluid intelligence scores are obtained from self-report questionnaires and test questions respectively, which are prone to intra-individual response instability, and potential social desirability bias in the case of neuroticism (McKelvie, 2004). These factors are reflected in previous estimates of test-retest reliability of neuroticism (0.85 in UKB (Dutt et al, 2021)) and fluid intelligence (0.65 in UKB (Lyall et al, 2016)). Response instability can be reduced by averaging over multiple available repeats of the same measure, thereby reducing noise and obtaining a more accurate estimate of the underlying trait.…”

“…In recent years, studies with sufficiently large participant samples (N>2000) have reported very low effect sizes for associations between neuroimaging measures and trait-level phenotypes (Dadi et al, 2021; Dutt et al, 2021; Marek et al, 2020). For example, Dadi et al used Random Forest Regression in N=10,000 participants from the UK Biobank (UKB) data to predict trait-level phenotypes of neuroticism and fluid intelligence, and reported a maximum R 2 of 0.04 when using neuroimaging features (Dadi et al, 2021).…”

“…For example, Dadi et al used Random Forest Regression in N=10,000 participants from the UK Biobank (UKB) data to predict trait-level phenotypes of neuroticism and fluid intelligence, and reported a maximum R 2 of 0.04 when using neuroimaging features (Dadi et al, 2021). Despite explaining a small amount of variance, results from these large participant samples are robust against sampling variability (Marek et al, 2022) and have good out-of-sample generalizability (Dadi et al, 2021; Dutt et al, 2021). Nevertheless, low brain-phenotype effect sizes are somewhat surprising given the presumed brain basis of phenotypic traits such as neuroticism and fluid intelligence.…”

“…The first potential explanation for low brain-phenotype effect sizes is that the neuroimaging features used for prediction are noisy measures of brain function. Previous work reported UKB test-retest reliability results for neuroimaging measures ranging from 0.3 for rfMRI measures to 0.9 for structural measures (Dutt et al, 2021). The relatively lower test-retest reliability of rfMRI measures likely results from both dynamic state fluctuations and measurement error due to the relatively small number of timepoints (t=490).…”