It has become increasingly apparent that the low reproducibility of results is a ubiquitous problem across many scientific fields such as biomedical science and psychology. This problem is particularly serious in biomedical studies using functional magnetic resonance imaging (fMRI) data. An increasing number of studies have reported success in constructing machine-learning algorithms (artificial intelligence) that use fMRI data to classify subjects as either healthy or suffering from a psychiatric disorder. However, it has been suggested that if these classifiers were constructed from a small number of samples (e.g., tens of participants) from a single site, it might not be possible to generalize their application to the data acquired from other imaging sites. A solution to this low generalization capability is to collect big data across many sites, but the considerable site-related differences in fMRI data is a formidable obstacle to the feasibility of this solution.
* This article was originally published here
