Self-learning classification of fMRI data with the CLASSIF1 software - an example with somatoform pain disorder

摘要

The aim of this study concerned the individual discrimination between patients with somatoform pain disorder and healthy controls by automated classification of cerebral fMRI activation patterns.This was achieved by applying a series of alternating noxious and innocuous tonic heat stimuli on the inner side of the left forearm to a group of female patients and gender and age matched healthy controls during fMRI scanning. The entire fMRI data time-series included noxious, innocuous and resting conditions. The data was classified following extraction and scaling of the time series of mean fMRI BOLD signals of 90 defined AAL ROIs with the SPM2 Marsbar Toolbox. The concatenated time-series of all ROIs for each patient was classified by the CLASSIF1 software (Valet,1993) using 10 randomly selected patients and 10 controls as learning set with 3 patients and 3 controls as unknown test set (hold-out validation).Classification accuracy was 92.3% (1 misclassification, each in the learning and test set). Discriminating differences of fMRI-BOLD signals between patients and controls were found in memory relevant brain structures, such as amygdala, fusiform cortex, parahippocampus, as well as the temporal cortex.CLASSIF1 classification allows accurate discrimination of individual patients from healthy controls with the new potential to elaborate interlaboratory standardized, hypothesis and model free fMRI data classifiers.