Emerging shifts in neuroimaging data analysis in the era of “big data”

Advances in positron emission tomography (PET) and functional magnetic resonance imaging (fMRI) have revolutionized our understanding of human cognition and its neurobiological basis. However, a modern imaging setup often costs several million dollars and requires highly trained technicians to conduct data acquisition. Brain-imaging studies are typically laborious in logistics and data management, and require costly-to-maintain infrastructure. The often small numbers of scanned participants per study have precluded the deployment of and potential benefits from advanced statistical methods in neuroimaging that tend to require more data (Bzdok and Yeo, NeuroImage 155:549–564, 2017; Efron and Hastie, Computer age statistical inference, 2016). In this chapter we discuss how the increased information granularity of burgeoning neuroimaging data repositories—in both number of participants and measured variables per participant—will motivate and require new statistical approaches in everyday data analysis. We put particular emphasis on the implications for the future of precision psychiatry, where brain-imaging has the potential to improve diagnosis, risk detection, and treatment choice by clinical-endpoint prediction in single patients. We argue that the statistical properties of approaches tailored for the data-rich setting promise improved clinical translation of empirically justified single-patient prediction in a fast, cost-effective, and pragmatic manner.