How does machine learning bias in prediction of behavioral phenotypes relate to individual characteristics?

Brain-based prediction of behavioral phenotypes is promising for precision medicine. However, predictive models in the field have commonly shown limited accuracy and individual bias. For instance, participants of ethnic/racial minority (e.g., African Americans) are more likely to be misclassified or show larger prediction errors, compared to white American participants. Addressing these limitations requires a better understanding of the sociodemographic and scan-related factors related to prediction biases. Here, we investigated the association between the prediction errors of brain-based models predicting over 100 behavioral measures and eight variables capturing sociodemographic and scan-related individual characteristics. These associations were examined in two developmental and one young adult cohorts (sample size = 4278, 432, and 784 respectively). We found that socioaffective and mental health phenotypes in developmental cohorts were particularly susceptible to bias associated with individual characteristics. In particular, head motion, head morphology, and sex/gender are consistently related to the machine learning biases in multiple behavioral measures. Our work highlights the multifactorial nature of the prediction errors generated by the brain-based predictive models. Overall, our results call for greater attention to model bias in developmental populations, particularly in the context of AI-based approaches for psychiatric risk prediction and early diagnosis.