New analysis of the software tools used to analyze functional magnetic resonance imaging (fMRI) data has discovered that the incidence of false positives (i.e. normal brain activity showing up on the scan as abnormal activity)  in one widely-used type of fMRI study is much higher than commonly believed. As a result, the chance of one or more false positives has been 70%, not the assumed 5%.

fMRI software relies heavily on statistical models, which in turn rely on certain assumptions about the behavior of the underlying data. One such assumption concerns spatial autocorrection, an fMRI phenomenon which can cause false positives. Those who built and validated the software determined a level of control for spatial autocorrection which kept the false positive rate at 5% in the dataset that was available to them. Unfortunately, that dataset was simulated, and when the authors of the newly published paper evaluated three common fMRI software tools using real data from healthy patients, they found that the level of control previously thought to be adequate was in fact seriously deficient. This led to a much-inflated false-positive rate.

These findings cast a long shadow, potentially affecting the results of thousands of papers that have used fMRI mapping studies, and more generally should serve as a reminder that even the oldest forms of neuroimaging are still not mature technologies. For more information about false-positive fMRI, see Neuroskeptic’s analysis of the paper.

 

 

Photo by OpenStax, CC BY 4.0.