Can machine learning-based predictive modelling improve our understanding of human cognition?

Abstract

A growing body of research predicts individual cognitive ability levels from brain characteristics including functional brain connectivity. The majority of this research achieves good prediction performance but provides limited insight into neurobiological processes underlying the predicted concepts. The insufficient identification of predictive characteristics may present an important factor critically contributing to this constraint. Here, we encourage to design predictive modelling studies with an emphasis on interpretability to enhance our conceptual understanding of human cognition. As an example, we investigated in a preregistered study which functional brain connections successfully predict general, crystallized, and fluid intelligence in a sample of 806 healthy adults (replication: N = 322). The choice of the predicted intelligence component as well as the task during which connectivity was measured proved crucial for better understanding intelligence at the neural level. Further, intelligence could be predicted not solely from one specific set of brain connections, but from various combinations of connections with system-wide locations. Such partially redundant, system-wide functional characteristics complement intelligence-relevant connectivity of brain regions proposed by established intelligence theories. In sum, our study showcases how future predictive studies on human cognition can enhance explanatory value by prioritizing a systematic evaluation of predictive characteristics over maximizing prediction performance. Significance StatementIntelligence represents a hallmark of human behavior, and a surge number of studies predicted individual scores from functional brain connectivity. However, actual understanding about its neural basis remains limited. We demonstrate how predictive modelling can be applied strategically to improve tracing predictive functional brain connections to enhance our understanding of intelligence. Our study unveils crucial findings about intelligence: differences in the neural code of distinct intelligence facets not detectable on a behavioral level and a brain-wide distribution of functional brain characteristics relevant to intelligence that extends those proposed by major intelligence theories. In a broader context, it offers a framework for future prediction studies that prioritize meaningful insights into the neural basis of complex human traits over predictive performance.