To err is human, but study shows mistakes may be predicted
Gwynne D. Koch
People performing routine, monotonous tasks occasionally are prone to making mistakes. According to Tom Eichele from the University of Bergen in Norway, examining the trial-by-trial dynamics in functional MRI activation patterns before, during and after an error could provide a better understanding of how performance monitoring works compared with just looking at the average activation after errors are made, as has been done in previous studies.
In functional MRI, stimulus-related activity is delayed and convoluted by the hemodynamic response, and the data typically are analyzed with a time series model. To examine the trial-by-trial dynamics, the scientists needed to recover the activity rather than to model it. This required a data acquisition and analysis technique that did not call for too many assumptions about the spatial or temporal features.
A study shows that MRI may be used to predict mistakes made while performing routine, monotonous tasks. Data analysis identified four independent components that predicted errors. The activation maps for these components are shown rendered onto transverse slices. Activation areas are depicted in red, deactivations in blue. Reprinted with permission of PNAS.
So Eichele and colleagues from the Medical Research Council Institute of Hearing Research in Southampton, UK, from the MIND Institute in Albuquerque, N.M., from Max Planck Institute in Cologne, Germany, and from other institutions employed a data-driven analysis technique to analyze patterns of brain activity that preceded errors on a trial-by-trial basis.
The team collected blood oxygen level-dependent response functional MRI data from participants who performed a forced-choice visual flanker task requiring rapid responses.
Principal component analysis was used to merge the data sets for all subjects so that they could be processed together, a required step for group independent component analysis, a technique that decomposes data into spatially nonoverlapping maps of regional activations, each with a representative activation time course.
The resulting output of spatial independent component analysis was an independent component map and an associated time course for every component and subject. This step was followed by deconvolution of the event-related hemodynamic response function for the independent components, from which the scientists could then estimate the amplitudes of the response in each trial. Using the Gift toolbox — free software that runs in Matlab — the analysis identified four independent components that reliably reflected brain activity modulated by task conditions and that collectively predicted errors.
In the flanker task, participants were instructed to press a button with the hand corresponding to the direction in which a central target arrow was pointing. The target was flanked by several arrows presented for 80 ms before the target, pointing in the opposite direction as the target in 50 percent of the trials. In these trials, a strong response tendency in the direction of the flanker arrows was created, leading to incorrect responses. Whenever participants were slow to respond, a symbolic feedback was presented, instructing them to speed up.
Imaging was performed at 3 T on a Siemens MRI system. Twenty-two functional slices were obtained using a gradient-echo echo-planar imaging sequence, with a resulting in-plane resolution of 3 × 3 mm. High-resolution anatomical and EPI-T1 images also were obtained.
The researchers found that a coincident increase in activity in the default mode network, a set of brain regions that underlies task-irrelevant mental activities, and a decrease in activity in frontal regions associated with maintaining task effort raised the probability of future errors. The temporal evolution of hemodynamic activation-predicted errors at least 6 s ahead in time — and as early as 30 s before a mistake was made — suggested that brain activity changes gradually toward an error-prone pattern while performing a routine task. However, once the subjects recognized their mistake or received feedback on late responses, the trend was interrupted and the activity pattern reset.
The scientists concluded that performance errors do not solely result from momentary fluctuations in concentration or brain activity, as has been previously suggested. They believe that their findings can inform models of performance monitoring, and they suggest that monitoring brain activity eventually could help predict — and possibly prevent — impending errors in critical real-world situations.
Next, the team will try to acquire similar data with concurrent EEG-functional MRI recordings, to tap the electrophysiological signal and its temporal resolution at the same time as the functional MRI signal.
PNAS, April 22, 2008, pp. 6173-6178.
MORE FROM PHOTONICS MEDIA