Volume-wise analysis of fMRI time series allows accurate prediction of language lateralization

Abstract

Using fMRI as a clinical tool, for example for lateralizing language, requires that it provides accurate results on the individual level. However, using a single voxel-wise activity map per patient limits how well the uncertainty associated with a decision can be estimated. Here, we explored how using a "volume-wise" analysis, where the lateralization of each time point of a patients fMRI session is evaluated independently, could support clinical decision making. Ninety-six patients with epilepsy who performed a language fMRI were analyzed retrospectively. Results from Wada testing were used as an indication of true language lateralization. Each patients 200 fMRI volumes were correlated with an independent template of prototypical lateralization. Depending on the strength of correlation with the template, each volume was classified as indicating either left-lateralized, bilateral or right-lateralized language. A decision about the patients language lateralization was then made based on how most volumes were classified. The results show that, using a simple majority vote, accuracies of 84% were reached in a sample of 63 patients with high-quality data. When 33 patients with datasets previously deemed inconclusive were added, the same accuracy was reached when more than 43% of a patients volumes were in agreement with each other. Increasing this cutoff to 51% volumes with agreeing classifications allowed for excluding all inconclusive cases and reaching accuracies over 90% for the remaining cases. Further increasing the cutoff to 65% agreeing volumes resulted in correct predictions for all remaining patients. The study confirms the usefulness of fMRI for language lateralization in patients with epilepsy, by demonstrating high accuracies. Furthermore, it illustrates how the diagnostic yield of individual volumes of fMRI data can be increased using simple similarity measures. The accuracy of our approach increased with the number of agreeing volumes, and thus allowed estimating the uncertainty associated with each individual diagnosis.