Reliability estimation of grouped functional imaging data using penalized maximum likelihood.

Abstract

We analyzed grouped fMRI data and developed a reliability analysis for such data using the method of penalized maximum likelihood (ML). Specifically, this technique was applied to a somatosensory paradigm that used a mechanical probe to provide noxious stimuli to the foot, and a paradigm consisting of four levels of graded peripheral neuromuscular electrical stimulation (NMES). In each case, reliability maps of activation were generated. Receiver operating characteristic (ROC) curves were constructed in the case of the graded NMES paradigm for each level of stimulation, which revealed an increase in the specificity of activation with increasing stimulation levels. In addition, penalized ML was used to determine whether the grouped reliability maps obtained from one stimulus level were significantly different from those obtained at other levels. The results show a significant difference (P < 0.01) in the reliability of activation from one stimulation level to the next. These results are in agreement with those obtained using generalized linear modeling (GLM). While the reliability maps generated are not directly comparable, they are qualitatively similar to those obtained by controlling the expected false discovery rate (FDR). The proposed methodology can be used to objectively compare activation maps between groups, as well as to perform reliability assessments. Furthermore, this method potentially can be used to assess the longitudinal effect of treatment therapies within a group.