Prediction of human errors by maladaptive changes in event-related brain networks

• deconvolution
• performance monitoring
• default mode
• frontal lobe

Human behavior may be strongly determined by the ongoing intrinsic dynamics of regional brain networks (1⇓–3). Although it is conceivable that patterns of brain activity preceding action execution are causally responsible for the behavioral outcome, previous research of human performance monitoring typically focused on brain activity evoked by and occurring after behavioral errors and the ensuing adaptive compensatory mechanisms (4⇓–6). It is not known yet how far ahead in time the state of cortical brain networks impacts on behavioral performance; some EEG and functional MRI (fMRI) evidence exists suggesting that activity in the preceding trial foreshadows subsequent errors but not earlier than that (7⇓⇓–10). The present study investigated this question by means of a single-trial analysis of event-related independent components (IC) in fMRI data during a visual task requiring rapid responses.

We analyzed patterns of evoked brain activity across trials that preceded erroneous behavior and may thus be instrumental in causing errors. Errors in this simple repetitive forced-choice task in a controlled setting may principally have two mutually nonexclusive causes: One source of errors consists simply of random failures to implement the correct behavior, which may be generated by momentary fluctuations in neural activity during stimulus processing and response selection in a given trial (1, 11). Although it is possible to make predictions about adaptive behavior in trials succeeding an error, it would be largely impossible in this condition to anticipate future errors from current brain states. Another possibility is that a proportion of errors may be caused by a systematic maladjustment in cognitive control systems that develops more slowly over time, and which thus affords a prediction about future behavior, in particular the likelihood of an erroneous response from analyzing trends in the history of activity across trials.

Models of cognitive control propose a performance-monitoring system involving the posterior medial frontal cortex (pMFC). This system tracks unfavorable outcomes and signals the need for appropriate behavioral adjustments (4, 6), which are thought to be initiated by enhancing and updating context and rule representations, thereby optimizing goal-directed behavior (12, 13). Notably, this performance-monitoring system signals the need for adjustment, which is required in response after errors and feedback and, more generally, whenever action outcome is at risk, for example, during response conflict (4). Maladjustment within cognitive control processes could build up gradually over a longer period and may be expressed as a decline of task-related attention and effort (14) and/or failure to suppress neural activity that interferes with performance.

An important source of activity that may interfere with behavioral performance is the so-called default mode network (DMN), a set of regions in the brain encompassing the precuneus, retrosplenial cortex, and anterior medial frontal cortex, which show increased metabolism during passive resting conditions (15, 16). These regions show correlated spontaneous fluctuations of hemodynamic activity among each other during relaxed wakefulness (17), that persist in sleep (18) and sedation (19). Also, anticorrelations between the DMN and regions relevant in diverse cognitive operations have been observed with functional connectivity analysis of fMRI data (2, 20, 21). The DMN is consistently extracted with other data-driven analysis techniques such as independent component analysis (ICA) of fMRI data recorded during relaxed resting (22), discrete sensory stimulation (23), continuous natural stimulation (24), and task performance (25). Additionally, the DMN has been linked to features in resting-state (26, 27) and event-related EEG (28). The range of functions that might involve the DMN and its interaction with other networks remains to be explored (29⇓–31). For our inquiry, it is important that, whereas DMN activity is typically suspended in the presence of cognitive tasks, parts of it are less deactivated to commission errors (10), and similarly reduced deactivation is found in stimulus-evoked hemodynamic responses during momentary lapses of attention, as defined by a slow response-time criterion in a visual attention task (11). Reduced activation in anterior cingulate and right prefrontal regions during the immediate prestimulus period has been associated with subsequent lapses (11), suggesting an antagonistic relationship between the DMN and areas controlling attention that is relevant for the outcome of goal-directed behavior.

Although Weissman et al. (11) focused primarily on the correlation between reaction time and activity patterns, the current analysis explores whether activity patterns that precede overt behavior foreshadow the outcome accuracy in some systematic way. We therefore assessed the relationship between errors in behavioral performance and the amplitude modulation of hemodynamic responses in error-preceding trials. In particular, we addressed the question how far ahead in time the state of brain networks impacts on behavioral accuracy. Such an analysis can provide important knowledge about the functionality of performance monitoring and causes of behavioral errors based on activity trends before the response and also yields measures that may aid in monitoring the likelihood of errors in real-world situations before a response is made.

In the flanker task we used [supporting information (SI) Fig. S1], errors are usually based on premature responses issued before stimulus analysis is completed (32). On incompatible trials, response conflict arises because the salient but distracting flankers and the target arrow drive two competing response tendencies. Errors occur if the flanker-induced response tendency is executed while the target-induced response tendency is still evolving (13). Performing correctly on trials involving high response conflict is cognitively demanding and engages a broad task-related network of brain regions, in particular the pMFC and lateral frontal, premotor, and anterior insular cortices and the intraparietal sulcus (12, 33, 34). Although functionally separable, these regions are overlapping with the proposed dorsal top-down attentional system (35), which has been reported to be anticorrelated to the DMN (21).

Here, we report on a set of brain regions in which activation patterns anticipate the accuracy of behavioral performance at least 6 sec ahead in time. To extract error-preceding activity, we analyzed data from participants who performed a speeded visual flanker task while blood oxygenation level-dependent response (BOLD) fMRI data were collected. In this task, participants had to respond to briefly presented arrows as quickly as possible. Trials consisted of either response-compatible or -incompatible flanker arrows inducing low- and high-response conflict, respectively. Performance errors and timing feedback resulted in behavioral adjustments such as speed/accuracy tradeoff. We have previously demonstrated the role of the pMFC in monitoring errors and influencing subsequent response time adjustments using an EEG-informed fMRI analysis of these data (36). Here, we focus on the identification of fMRI signals that systematically predict errors. This is challenging, because it is usually not possible to specify a model regarding the actual timing and modulation of the underlying precursor signals and adequate parameters of the associated hemodynamic response function (HRF). Therefore, we used a data-driven approach and decomposed the fMRI data with spatial ICA (37, 38). After ICA, we deconvolved HRFs from the timecourses of individual event-related components (ICs) to recover individually and regionally specific empirical estimates of the HRF (39, 40) and used these subsequently for estimation of single-trial responses. Thereby, we effectively removed the effect of hemodynamic convolution and recovered the trial-by-trial amplitude modulation of event-related responses (Fig. 2). With this analysis, four ICs were identified that reliably reflected brain activity modulated by task conditions and that collectively predicted errors (Fig. 1 and Table S1). In this set of brain regions, the temporal evolution of hemodynamic activation-predicted performance errors at least 6 sec ahead in time, with linear trends starting as early as 30 sec before an erroneous response. A coincident decrease of deactivation in regions of the DMN, together with a decline of activation in frontal regions associated with resource allocation and maintaining task effort, raised the probability of future errors.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Component maps, hemodynamic responses, and functional modulation. a, d, g, and j show the activation maps of the four ICs rendered onto the MNI template at representative transverse (Left), coronal (Center), and sagittal (Right) slices, coordinates in millimeters are given at the top-right corners. The maps are shown in neurological convention (left hemisphere is on the left). IC areas were considered significant when exceeding a cluster extent of at least 27 contiguous voxels, and at 1% false-positive discovery rate, thresholds are given below each IC map. Activations are plotted in red, deactivations in blue. b, c, e, f, h, i, k, and l display the HRFs within the respective ICs as estimated via deconvolution from 1 to 20 sec after stimulus onset, in arbitrary units. The group average from the 13 participants is plotted as a solid line; error bars indicate ± 1 SEM, and dots represent all individual estimates. The empirical HRFs were used to estimate single-trial amplitudes in the fMRI data to assess systematic preerror activity, here group averaged (±1 SEM) trial sequences from −6 to +2 around response errors are shown (c, f, i, and l). These plots represent the residual data after removal of variability accounted for by response conflict, feedback, stimulus sequence, and response times.

Materials and Methods

We analyzed data from 13 healthy right-handed participants (22–29 years, eight female) who performed a speeded flanker task while fMRI data were recorded. Written informed consent was obtained from each participant before the start of the experiment according to the declaration of Helsinki. Details about the flanker task and relevant behavioral effects are provided in Figs. S1–S3.

Image Acquisition.

Imaging was performed at 3 T on a Siemens Trio system equipped with the standard bird cage head coil. Twenty-two functional slices were obtained parallel to the anterior commissure–posterior commissure line (thickness, 4 mm; interslice gap, 1 mm) using a gradient-echo echo planar imaging (EPI) sequence with an echo time of 30 ms, a flip angle of 90°, a repetition time of 2,000 ms, and an acquisition time of 1,500 ms. The fMRI matrix acquired was 64 × 64 with a field of view of 19.2 cm, resulting in an in-plane resolution of 3 × 3 mm². A total of 1,309 volumes were acquired. Before the functional runs, anatomical modified driven equilibrium Fourier transform, and EPI-T1 slices in the plane with functional images were collected. The average intertrial interval amounted to 6 sec, and trial onsets were jittered ± 1.5 sec relative to volume acquisition to improve temporal sampling of the BOLD response.

Preprocessing.

All images were realigned to the first image in the time series to correct for head movement and then normalized to the MNI reference space using SPM2 (www.fil.ion.ucl.ac.uk/spm). Normalized data were resliced to a cubic voxel size of 2 mm³ and smoothed with a Gaussian kernel with 8 mm FWHM.

ICA.

We used the rationale for group ICA proposed by Calhoun et al. (38, 50) implemented in the GIFT toolbox (http://icatb.sourceforge.net) and extended it with deconvolution and single-trial estimation. For clarity, we provide a detailed description of the group ICA model in SI Text, complementing the graphical illustration in Fig. 2. Group ICA requires that all subjects are analyzed at once, and principal component analysis (PCA) was used for compression to allow the datasets to be processed together (50). In the PCA step, data from each subject were reduced from the number of time points within the experiment (n = 1,306) to 52 dimensions. We adjusted the number of dimensions that were estimated in this study heuristically to accommodate extraction of the maximum number of components that was computationally feasible and yield a high replicability of ICs across runs while consistently separating sources representing artifacts and known components of interest (22). In our experience, the exact choice of the number of components does not critically affect the results as long as this number is not smaller than the true dimension. Spatial ICA was then performed by using the infomax algorithm (51), with subsequent back reconstruction into single subjects. The resulting output is an independent component map and an associated IC timecourse for every component and subject, which were subjected to random effects analyses. Replicable ICs were inspected to identify and discard those primarily associated with artifacts colocalizing with gray matter (e.g., motion, flow, susceptibility; see Fig. 2 map-based criteria). From the remaining ICs, only those with significant random-effects t statistics of their maps, adjusted for multiple comparisons at 1% false-positive discovery rate and cluster extent of at least 27 contiguous voxels were considered further. For these ICs, the individual timecourses were filtered with a 72-sec high-pass fifth-order Butterworth digital filter, averaged across replications and normalized to unit variance.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Schematic of the group ICA analysis with deconvolution and single trial estimation. In the group ICA model, we assume that the fMRI is a linear mixture of spatially independent sources. The mixing of sources is represented by A and yields the ideal samples of brain activity u_p(v_p) and the signals recorded with functional magnetic resonance imaging (B). Transformations (T) during preprocessing change the sampling properties. For each individual, the preprocessed data were reduced to F via PCA. Group data are generated by concatenating individual principal components in G. Spatial ICA was performed in this set, estimating aggregate components (C). Individual maps and timecourses were then back-reconstructed and inspected to identify and discard those associated with artifacts. From the remaining ICs, only those with significant maps were considered further. Then, the HRF was deconvolved by forming the convolution matrix of the stimulus onsets and multiplying the pseudoinverse with the IC timecourse. Single-trial estimation was done by fitting a design matrix (X) containing predictors for the onset times of each trial convolved with the estimated HRF onto the IC timecourse (y), estimating the scaling coefficients (β) of the regression model y = β·X + ε. See SI Text for further detail.

Deconvolution.

The event-related HRF for the ICs was then deconvolved by forming the convolution matrix of the stimulus onsets with an assumed kernel length of 20 sec and multiplying the pseudoinverse of this matrix with the IC timecourse. Single-trial estimation was performed only in ICs where consistent event-related HRFs were present, i.e., where yielding a mono-/biphasic form with a peak latency between 3 and 12 sec in all subjects. Single-trial response amplitudes were then recovered by fitting a design matrix (X) containing separate predictors for the onset times of each trial convolved with the estimated HRF onto the IC timecourse, estimating the scaling coefficients (β) in the multiple linear regression model y = β·X + ε using least squares. Single-trial β estimates were subsequently entered into within-subjects one-sample t tests, and only those with significant (P < 0.001) differences from zero magnitude in all participants were considered further.

Inference.

The last step in component selection was to identify the functionally relevant ICs. First, we modeled the respective single-trial weights of each subject with multiple linear regression including predictors coding for the trial type (compatible, incompatible), trial sequence (same, different), outcome (correct, error), timing feedback (“speed up,” absent), response times, and the trial-to-trial changes of response times. Correspondingly, we tested for any significant difference between the component mean responses to the different trial categories by entering averaged weights for compatible and incompatible correct trials divided into fast and slow responses (mean split), feedback, and errors into an ANOVA with repeated measures. The subset of ICs reported here showed significant event-related responses in general; consistent effects related to response conflict, accuracy, and feedback; and differences between the conditions (Fig. S4).

Four additional ICs localizing to the visual system, the dorsal and ventral attentional system (35), and the anterior part of the DMN, respectively, showed robust event-related HRFs but did not appear significantly modulated by the current task and were thus not considered further for error precursor activity (Fig. S5).

In the selected components, we removed the variability associated with the aforementioned predictors with multiple regression and used the residual data to check for the presence of error-preceding activity. This was done by generating averages from the residuals for sequences from six trials before to three trials after error responses, that is, from ≈30 to 40 sec preceding the event to 15–20 sec after it. The individual sequences were then entered into pointwise one-sample t tests under the assumption of zero magnitude. IC 3 and 4 were additionally tested for a linear trend expanding from trial position T₋₆ through T₋₁ preceding an error using linear regression. Individual β-estimates were entered into one-sample t tests, and effects were considered significant at P < 0.05.

Additionally, we assessed the connectivity among all four components by computing pairwise correlation coefficients between the single-trial estimates in each participant and subjected the coefficients to one-sample t tests after Fisher transformation. Results are reported in Table S2.

To estimate the predictive value of the four identified ICs for future behavior, we computed the change in error frequency relative to the base rate (absolute error rate 8.7%) for current trials (T₀) using a linear fit across the residual activities of the four ICs in the sequence of the preceding five trials (T₋₆ to T₋₁) for all single-trial sequences pooled across all participants. Based on a mean split of the beta weights of the linear fit in the four ICs, we generated 16 nonoverlapping classes of trials for which we computed the error frequency at T₀ and the proportion of feedback trials, Compatible trials and average residual response time (after removal of categorical effects from the raw response times) at T₋₁ to control for these possible confounds. Results are reported in Table S3.

• Received September 21, 2007.

• © 2008 by The National Academy of Sciences of the USA