Disentangling working memory processes during spatial span assessment: a modeling analysis of preferred eye movement strategies

Abstract

The neurocognitive processes involved during classic spatial working memory (SWM) assessment were investigated by examining naturally preferred eye movement strategies. Cognitively healthy adult volunteers were tested in a computerized version of the Corsi Block-Tapping Task--a spatial span task requiring the short term maintenance of a series of locations presented in a specific order--coupled with eye tracking. Modeling analysis was developed to characterize eye-tracking patterns across all task phases, including encoding, retention, and recall. Results revealed a natural preference for local gaze maintenance during both encoding and retention, with fewer than 40% fixated targets. These findings contrasted with the stimulus retracing pattern expected during recall as a result of task demands, with 80% fixated targets. Along with participants' self-reported strategies of mentally "making shapes," these results suggest the involvement of covert attention shifts and higher order cognitive Gestalt processes during spatial span tasks, challenging instrument validity as a single measure of SWM storage capacity.

Figure 1

Time sequence for a typical trial of the spatial span paradigm, divided into three phases: 1) Encoding, where ten circles were displayed on the computer screen, and a subset of 4, 6, or 8 circles changed serially from white to gray in a specific order; 2) Retention, where the white background circles remained displayed for durations of 5 or 20 seconds, and 3) Recall, where participants were instructed to click on the targets in the same order that they were presented. The onset of recall was indicated by the appearance on the screen of a blue box, a “Done”-button, and a cursor.

Figure 2

Example of construction of the ON-TARGET and ZONE models for a stimulus where 6 targets had to be remembered. The likelihood function of the ON-TARGET model (top panel) was characterized by 6 two-dimensional Gaussians with identical parameterized standard deviation (σ) built above the to-be-remembered targets (T1–T6); and by a multinomial coefficient representing the evenness of the fixation distribution with respect to a target-based Voronoi space decomposition. The ZONE model (bottom panel) was characterized by a unique two-dimensional Gaussian with parameterized central coordinates (x_c, y_c) and standard deviation (σ).

Figure 3

Hypothetical datasets generated in four simulations to test the good functioning of the ON-TARGET and ZONE models. These simulations assumed a randomly generated stimulus configuration where 6 targets, T1 to T6, had to be remembered. Hypothetical datasets were generated assuming 30 virtual participants, each performing a random number of fixations between 1 and 30. Fixation locations were randomly selected within specific two-dimensional distributions, defined by 6 two-dimensional Gaussians (σ_sim=75 pixels) centered on targets T1 to T6 for Simulation 1; by a unique Gaussian (σ_sim=75 pixels) centered on the middle of the screen for Simulations 2; by a unique Gaussian (σ_sim=75 pixels) centered on target T3 for Simulation 3; and by a flat distribution, covering the screen evenly for Simulation 4. Maximum likelihood parameters resulting from each model’s best fit are illustrated in the form of circles of radius the estimated standard deviations (σ): black dotted circles centered on targets T1 to T6 for the ON-TARGET model, and gray dotted circles with estimated (x,y) centers for the ZONE model.

Figure 4

Fit results for the four simulated datasets generated to test the good functioning of the ON-TARGET and ZONE models. For each simulation, Log-Likelihood functions are presented for the ON-TARGET model (solid line) and ZONE model (dashed line) with dependence on the Gaussian’s standard deviation parameter (σ).

Figure 5

Descriptive modeling results for the fixation distribution data, calculated separately for each of the 24 stimuli (x-axis) and paradigm phase (encoding, retention, and recall). Modeling results are presented for the ON-TARGET (black squares) and ZONE (light gray diamonds) models in terms of maximum log-likelihood, a measure of goodness of fit (upper panels), and in terms of estimated parameter σ_ML, i.e., the Gaussians’ standard deviation at maximum likelihood (lower panels). The 24 stimuli, which were presented in a different order during paradigm administration for each participant, are re-ordered on the graphs based on stimulus characteristics, i.e., target number and delay duration.

Figure 6

Fixation distribution data obtained during encoding, retention, and recall for 3 representative stimuli, where 4, 6, and 8 targets had to be remembered and with a retention delay of 5 seconds. Data are presented in the form of heat maps with darker color corresponding to higher fixation density. Model fit parameter results are illustrated on the graphs in the form of circles of radius the estimated Gaussians’ standard deviation at maximum likelihood (σ_ML): black dotted target-centered circles for the ON-TARGET model and gray solid circles with centers estimated by model fit for the ZONE model.

Figure 7

Goodness of fit modeling results presented in terms of Bayesian Information Criterion (BIC) for fixation distributions recorded during encoding, retention, and recall. Combined BIC scores are provided by summing maximum log-likelihood values for stimuli with same target number, assuming data independence across trials after accounting for eye movement strategy. BIC results are presented separately for the ON-TARGET (black) and ZONE (gray) models, and model fit differences (ΔBIC) are presented in gray bar-graphs.

Figure 8

Proportion (upper panel) and number (lower panel) of targets fixated by participants during encoding, retention, and recall, calculated separately for each of the 24 stimuli (x-axis). The 24 stimuli are ordered in the graphs based on stimulus characteristics, i.e., target number and delay duration.

Figure 9

Proportion of targets fixated for trials with delays of 5 seconds (left panel) or 20 seconds (right panel), combining trials with same target number. Differences between encoding, retention, and recall that reached significance per the result of the repeated measure MANOVAs are indicated with asterisks.