Gast ::
| Hauptseite > Publikationsdatenbank > Cross-ethnicity/race generalization failure of RSFC-based behavioral prediction and potential downstream consequences |
| Hauptseite > Publikationsdatenbank > Cross-ethnicity/race generalization failure of RSFC-based behavioral prediction and potential downstream consequences |
| Conference Presentation (After Call) | FZJ-2022-02752 |
2022
Abstract: Algorithmic biases that favor majority populations pose a key challenge to the application of machinelearning for precision medicine. In neuroimaging, there is growing interest in the prediction ofbehavioral phenotypes based on resting-state functional connectivity (RSFC). In that context, predictivemodels are typically built by capitalizing on large cohorts with mixed ethnic groups, in which theproportions of certain groups, e.g. African Americans (AA), are limited. Here, we investigated cross-ethnicity/race generalizability of the current, field-standard behavioral prediction approach using twolarge-scale public datasets from the United States. Specifically, we observed larger prediction errors inAA than white Americans (WA) for most behavioral measures using both the Human ConnectomeProject (HCP) and the Adolescent Brain Cognitive Development (ABCD) data. This prediction biastowards WA corresponded to more WA-like brain-behavior association patterns learned by the models.Looking into the direction of prediction errors, concerns can be raised if the machine-learning predictionresults would be uncritically used, in particular for the diagnosis of mental disorders. For example, socialsupport measures were more overpredicted for AA than WA, whereas social distress measures such asPerceived Rejection were more underpredicted for AA than WA.Furthermore, African pre-adolescent participants suffered from more overpredicted social problems,rule-breaking and aggressive behaviors compared to white participants. Effects of the trainingpopulations were also studied by comparing predictive models trained specifically on AA, specifically onWA, or on a mixture of AA and WA with equal sizes. Although specific training on AA slightly helped toreduce the biases against AA, most behavioral measures still exhibited larger prediction errors in AAthan WA. Other possible sources of the biases such as neuroimaging preprocessing (e.g., brain templatesand functional atlases) and the design of behavioral measures need to be examined in the future.
|
The record appears in these collections: |
