The Cost of Untracked Diversity in Brain-Imaging Prediction

Abstract

Brain-imaging research enjoys increasing adoption of supervised machine learning for single-subject disease classification. Yet, the success of these algorithms likely depends on population diversity, including demographic differences and other factors that may be outside of primary scientific interest. Here, we capitalize on propensity scores as a composite confound index to quantify diversity due to major sources of population stratification. We delineate the impact of population heterogeneity on the predictive accuracy and pattern stability in two separate clinical cohorts: the Autism Brain Imaging Data Exchange (ABIDE, n=297) and the Healthy Brain Network (HBN, n=551). Across various analysis scenarios, our results uncover the extent to which cross-validated prediction performances are interlocked with diversity. The instability of extracted brain patterns attributable to diversity is located preferentially to the default mode network. Our collective findings highlight the limitations of prevailing deconfounding practices in mitigating the full consequences of population diversity.

Connectome-based machine learning models are vulnerable to subtle data manipulations

“…The neuroimaging community is beginning to recognize and explore the impacts of ethics in machine learning models, with a particular focus on bias in datasets and models (Benkarim et al 2021). Trust is distinct from bias, and it is an equally important yet widely overlooked facet of ethics in neuroimaging models.…”

“…Connectome-based predictive models are at the forefront of this trend (Finn and Rosenberg 2021;Shen et al 2017), showing promising results in understanding general cognition (Beaty et al 2018;Song et al 2021;Dubois et al 2018;Rosenberg et al 2018) and mental health (Du et al 2018;Lynall et al 2010;Nielsen et al 2020). Improvements in accuracy (Cui and Gong 2018;Gan et al 2021;Li et al 2021;Kohoutová et al 2020) and fairness (i.e., lack of bias (Benkarim et al 2021)) of connectome-based models represent an important step in preparing these models for real-world applications. But, accurate and bias-free models are not enough.…”

Brain–phenotype models fail for individuals who defy sample stereotypes

“…Together, this reflects a paradigm shift in human neuroscience research from a focus on the group to a focus on the individual, with important potential applications to clinical practice [21][22][23] .To deliver on this promise, however, these approaches must identify patterns of brain activity that are relevant to the phenotype of interest in a given individual-the patient sitting before their clinician, for example. Previous linear modelling work has relied on the assumptions that (1) a single brain network is associated with a given phenotype, with patterns of activity within that network varying across individuals 10,24,25 ; and (2) larger, more heterogeneous samples will more accurately and reliably capture this single model 26,27 . But although many published models have demonstrated impressive generalizability 6,9,10 , they do not account for brain-phenotype relationships in all individuals 13,14 .…”

“…But although many published models have demonstrated impressive generalizability 6,9,10 , they do not account for brain-phenotype relationships in all individuals 13,14 . This raises the crucial question of in whom models fail, and why.The existence of structured model failure-some individuals who are better fit by a model than others 14,24,26 -would suggest that one brain-phenotype relationship does not fit all, and that systematic bias may determine who is fit and who is not. This, in turn, may engender imprecise, misleading and in some cases harmful model interpretations.…”

“…This, in turn, may engender imprecise, misleading and in some cases harmful model interpretations. That is, a brain network that is found to be associated with a given phenotype may only apply to a specific subset of the population at large, limiting its practical utility 14,26,28 , or may not represent the phenotype of interest. Indeed, factors that interfere with adequate phenotypic characterization have been documented for many widely used neurocognitive tests 18,29 , and may include the fallacy of universalism (construct bias), the application of inappropriate norms, discordance between primary and assessment language and the presence of…”