Single trial MEG recordings can predict the subject's ability to recognize a natural scene

• The mask was followed by a confidence rating screen. By moving the finger on the side of the green triangle subjects indicated a high confidence in their ability to recognize the target picture. • The query phase immediately followed the confidence rating. The target was presented together with three distractor pictures (4 AFC). The subject had to indicate the target image by finger movement. • Every picture was presented only once. The mask consisted of randomly sized and oriented rectangles and lines, mimicking the statistical properties of natural images. The colors for the mask were randomly chosen from all four pictures shown on that trial. • Magnetic fields were recorded with a CTF 151-channel MEG-whole head system while target and mask were presented. • Seven subjects participated in the experiment. • We only included trials in which the confidence rating matched the actual response, in order to avoid lucky guesses and careless errors. • Data recorded within 600 ms from target onset were included in the classification. • Trials were grouped into correct and false trials. For each channel and each point in time t-values were calculated between these two groups. • Data reduction was achieved by applying a region growing algorithm to cluster the data. Two parameters were specified: First only samples above a selected t-value were included. The second parameter was the minimum number of spatial and temporal neighbors that must fulfill the t-criterion. Contiguous samples that satisfy these two criteria are then combined into a cluster. The cluster means were then included into the further analysis • The means of the clusters from single trials were classified in a leave-one-out cross-validation , e.g. the models for the classification were calculated from all but the current trial. • Mean and variance of the cluster means over trials were estimated. For the cluster data of the given trial two z-values were calculated from these parameters, one for the correct and one for the false trials. • The trial was assigned to the class in which the sum of the squared z-values was lowest (χ 2). • On average 79 percent of the trials were correctly classified. • Confidence limits for the classification performance were determined in a randomization test. Classification performance obtained with the measured datasets was significantly above chance for all subjects (t 2 and neighbours 2). ¾ It is possible to predict with reasonable performance (83% …