An Eye Fixation-Related Potential Study in Two Reading Tasks: Reading to Memorize and Reading to Make a Decision

Abstract

We investigated how two different reading tasks, namely reading to memorize [Read & Memorize (RM)] and reading to decide whether a text was relevant to a given topic [Read & Decide (RD)], modulated both eye movements (EM) and brain activity. To this end, we set up an ecological paradigm using the eye fixation-related potentials (EFRP) technique, in which participants freely moved their eyes to process short paragraphs, while their electroencephalography (EEG) activity was recorded in synchronization with their EM. A general linear model was used to estimate at best EFRP, taking account of the overlap between adjacent potentials, and more precisely with the potential elicited at text onset, as well as saccadic potentials. Our results showed that EM patterns were top-down modulated by different task demands. More interestingly, in both tasks, we observed slow-wave potentials that gradually increased across the first eye fixations. These slow waves were larger in the RD task than in the RM task, specifically over the left hemisphere. These results suggest that the decision-making process during reading in the RD task engendered a greater memory load in working memory than that generated in a classic reading task. The significance of these findings is discussed in the light of recent theories and models of working memory processing.

Appendices

Appendix 1: EFRP by Averaging with an Individual Baseline

Results from the ANOVA revealed the following significant effects:

Condition: F(1, 19) = 23.16, p < .001, ηp² = .55

Condition × Hemisphere interaction: F(1, 19) = 4.74, p < .05, ηp² = .20

Condition × Hemisphere × ROI interaction: F(2, 38) = 7.62, p = .01, ηp² = .29

Condition × Hemisphere × Rank interaction: F(3, 57) = 3.20, p < .05, ηp² = .14

Condition × ROI × Rank interaction: F(6, 114) = 3.56, p < .05, ηp² = .16

Condition × Hemisphere × ROI × Rank interaction: F(6, 114) = 2.98, p < .05, ηp² = .14

Table 3 sets out the results of the post hoc comparison for the 4-factor interaction (i.e., Condition × Hemisphere × ROI × Rank), for frontal, central and parietal ROI. Figure 9 illustrates the EFRP results with an individual baseline for the frontal ROI (i.e., F4, F6, F8 and FC6 for the right hemisphere and F3, F5, F7 and FC5 for the left one).

Table 3 Mean (standard error) amplitude values, in µV, in the latency ranges of interest for frontal, central and parietal ROI in both conditions with an individual baseline

Fig. 9

Grand-average EFRP, estimated by averaging with an individual baseline, for the RM task (top) and RD task (bottom), for the first four fixations on the text, for the frontal right (right) and left (left) ROI

As emphasized in Subsection 3.3.2 (“Estimation by averaging”), an individual baseline resets the activity to zero (Körner et al. 2014) at each fixation onset. For example, the mean activity of the EFRP baseline at fixation n is roughly equivalent to the mean activity elicited at fixation n − 1. Our results showed no major increment or decrement in synchronized activity elicited at one fixation compared with the following fixation, showing that the level of synchronized activity at fixation rank n was roughly the same as that at fixation rank n + 1. As expected, slow-wave activity was not synchronized with each fixation. There were several significant differences between the two tasks, with larger mean amplitudes for the RD task than for the RM one, showing that the contribution of a specific fixation to slow-wave amplitude was greater in the RD condition than in the RM one.

Appendix 2: Implementation and Configuration of the GLM

In the present study, GLM were used to estimate the neural activity time-locked to the period between first fixation onset and fourth fixation offset. This activity can be written as $fp(t)$. The latency of the first fixation onset was around 200 ms (see Table 4), showing that, as previously emphasized, this activity overlapped with the potential elicited at text onset $s(t)$.

Table 4 Statistical summary of the latency (in ms) of the first and fourth fixation onsets based on individual means

These two potentials were included in the model as described by Eq. 3 (see Subsection 3.4.1), with the potential evoked by the first fixation plus the potentials elicited by the successive fixations in the selected time window. The saccadic potentials elicited at the four first saccade onsets were also considered in the model, as the distribution of the incoming saccades, in terms of direction and amplitude, differed across conditions (see Subsection 3.2.2). Let us recall here the equation ${x_i}\left( t \right)=s\left( t \right)+~fp(t - \tau _{i}^{{(1)}})+\mathop \sum \limits_{{c=1}}^{4} \mathop \sum \limits_{l} s{p_c}(t - \tau _{{c,i}}^{{'\left( l \right)}})+{n_i}(t)$, where ${x_i}\left( t \right)$ is the signal time-locked to text onset during the $ith$ epoch, $s(t)$ the potential elicited at text onset, $fp(t)$ the potential evoked at first fixation onset plus the potentials elicited at subsequent fixation onsets up to the end of the epoch, $s{p_c}\left( t \right)$ the saccadic potential elicited at saccade onset for a given category ($c$), $\tau _{i}^{{(1)}}$ the timestamp of the first fixation onset, $\tau _{{c,i}}^{{'\left( l \right)}}$ the timestamp of the lth saccade onset for the category $c$, and ${n_i}(t)$ the noise of the ongoing activity. We considered four categories of saccades: progressive vs. regressive, and short vs. long.

This appendix shows the mathematical implementation leading to the final estimation. To estimate $s(t)$ and $f_p(t)$ by ordinary least square regression, Eq. 3 can be rewritten in matrix form:

$$\forall i \in \left\{ {1, \ldots ,E} \right\},~~{{\varvec{x}}_{\varvec{i}}}={{\varvec{D}}_s}.{\varvec{s}}+{{\varvec{D}}_{fp,i}}~.{\varvec{f}}{\varvec{p}}+\mathop \sum \limits_{{c=1}}^{4} {{\varvec{D}}_{sp,c,i}}.{\varvec{s}}{{\varvec{p}}_{\varvec{c}}}+{{\varvec{n}}_i}.$$

(4)

${{\varvec{x}}_i}$ is the vector (${{\varvec{x}}_i}={\left[ {{x_i}\left( 1 \right),~ \ldots ,~{x_i}\left( {{N_e}} \right)} \right]^\dag };~{x_i} \in {{\mathbb{R}}^{{N_e}}}$) of the observed EEG samples time-locked to text onset, for the $ith$ epoch, with ${[.]^\dag }$ the transpose operator. ${N_e}$ is the number of samples, that is, the length of the observed signal ${x_i}\left( t \right)$. ${{\varvec{n}}_i}$ is the noise vector (${{\varvec{n}}_i}={\left[ {{n_i}\left( 1 \right),~ \ldots ,~{n_i}\left( {{N_e}} \right)} \right]^\dag };~{{\varvec{n}}_i} \in {{\mathbb{R}}^{{N_e}}}$) with the same number of samples.$~{\varvec{s}} \in {{\mathbb{R}}^{{N_s}}}$ is the vector of the response time-locked to text onset, and ${N_s}$ is the length of the response $s(t)$. ${\varvec{f}}{\varvec{p}} \in {{\mathbb{R}}^{{N_{fp}}}}$ is the vector of the response time-locked to first fixation onset and ${N_{fp}}$ is the length of the response $fp(t)$. For a given saccade category $c$, ${\varvec{s}}{{\varvec{p}}_{\varvec{c}}} \in {{\mathbb{R}}^{{N_{sp}}}}$ is the vector of the saccadic response time-locked to saccade onset and ${N_{sp}}$ is the length of the response $s{p_c}(t)$. ${{\varvec{D}}_s} \in {{\mathbb{R}}^{{N_e} \times {N_{\text{s}}}}}$ is the Toeplitz matrix⁵ with ${N_e}$ rows and ${N_s}$ columns, for text onset. ${{\varvec{D}}_s}$ is defined by its first column, with entries that are all equal to zero except for one at the row subscript corresponding to the $0$ ms position in the epoch. ${{\varvec{D}}_{fp,i}} \in {{\mathbb{R}}^{{N_e} \times {N_{fp}}}}$ is the Toeplitz matrix with ${N_e}$ rows and ${N_{fp}}$ columns, for coding the first fixation onset during the $ith$ epoch. ${{\varvec{D}}_{fp,i}}$ is defined by its first column, with entries that are all equal to zero except one, at the row subscript corresponding to the $\tau _{i}^{{(1)}}$ ms position in the $ith$ epoch. The Toeplitz matrices ${{\varvec{D}}_{sp,c,i}}$ ($c=1~..~4$), ${N_e}$ rows and ${N_{sp}}$ columns, were set up in the same way, but using the timestamps $\tau _{{c,i}}^{{'\left( l \right)}}$ for the row subscripts. Unlike the matrices ${{\varvec{D}}_{fp,i}}$ and ${{\varvec{D}}_{sp,c,i}}$, the ${{\varvec{D}}_s}$ matrix does not depend on the epoch number ($i$), as the timestamps of text onset are always equal to zero whatever the epoch. ${{\varvec{D}}_s}$, ${{\varvec{D}}_{fp,i}}$ and ${{\varvec{D}}_{sp,c,i}}$ ($c=1..4$) are sparse matrices. Matrices ${{\varvec{D}}_s}$ and ${{\varvec{D}}_{fp,i}}$ are composed of only one diagonal equal to one, all other values being equal to zero. For the matrices ${{\varvec{D}}_{sp,c,i}}$ ($c=1..4$), zero, one, two, three or four diagonals are equal to one, depending on the number of saccades belonging to a given category ($c$), and all other values are equal to zero. Considering all epochs ($E$), the observations are concatenated such that:

$${\varvec{x}}={{\varvec{D}}_S}.{\varvec{s}}+{{\varvec{D}}_{Fp}}.{\varvec{f}}{\varvec{p}}+\mathop \sum \limits_{{c=1}}^{4} {{\varvec{D}}_{Sp,c}}.{\varvec{s}}{{\varvec{p}}_{\varvec{c}}}+{\varvec{n}}$$

(5)

with ${\varvec{x}}={\left[ {{\varvec{x}}_{1}^{\dag },{\text{~}} \ldots ,{\text{~}}{\varvec{x}}_{E}^{\dag }} \right]^\dag } \in {{\mathbb{R}}^N}$, ${{\varvec{D}}_S}={\left[ {{\varvec{D}}_{s}^{\dag }, \ldots ,{\varvec{D}}_{s}^{\dag }} \right]^\dag } \in {{\mathbb{R}}^{N \times {N_s}}}$, ${{\varvec{D}}_{Fp}}={\left[ {{\varvec{D}}_{{fp,1}}^{\dag }, \ldots ,{\varvec{D}}_{{fp,E}}^{\dag }} \right]^\dag } \in {{\mathbb{R}}^{N \times {N_f}}}$ and ${{\varvec{D}}_{Sp,c}}={\left[ {{\varvec{D}}_{{sp,c,1}}^{\dag }, \ldots ,{\varvec{D}}_{{sp,c,E}}^{\dag }} \right]^\dag } \in {{\mathbb{R}}^{N \times {N_{sp}}}}$ ($c=1..4$), where $N=~{N_e} \times E$, $N$ is the total number of samples. After concatenation of the Toeplitz matrices and evoked potentials, Eq. 6 becomes ${\varvec{x}}={\varvec{D}}.{\varvec{p}}+{\varvec{n}}$ with ${\varvec{D}}$ equal to $[{{\varvec{D}}_S},{{\varvec{D}}_{Fp}},~{{\varvec{D}}_{{\varvec{S}}{\varvec{p}},1}},~..,~{{\varvec{D}}_{{\varvec{S}}{\varvec{p}},4}}]$ and ${\varvec{p}}~$is the concatenation of the six evoked potentials such that ${\varvec{p}}={\left[ {{{\varvec{s}}^\dag },{\varvec{f}}{{\varvec{p}}^\dag },{\varvec{s}}{\varvec{p}}_{1}^{\dag },~ \ldots ,~{\varvec{s}}{\varvec{p}}_{4}^{\dag }~} \right]^\dag }$. The solution given by the least square minimization is:

$${\widehat {{\varvec{p}}}_{GLM}}={\left( {{{\varvec{D}}^\dag }.{\varvec{D}}} \right)^{ - 1}}.{\varvec{D}}.{\varvec{x}}$$

(6)

where ${\widehat {{\varvec{p}}}_{GLM}}$ is the concatenation of both estimates, such that ${\widehat {{\varvec{p}}}_{GLM}}={\left[ {{{\widehat {{\varvec{s}}}}_{GLM}}^{\dag },{{\widehat {{{\varvec{f}}{\varvec{p}}}}}_{GLM}}^{\dag },{{\widehat {{{\varvec{s}}{\varvec{p}}}}}_1}{{_{{GLM}}}^\dag },~..,~{{\widehat {{{\varvec{s}}{\varvec{p}}}}}_4}{{_{{GLM}}}^\dag }} \right]^\dag }.$

As for the estimation by averaging, the model was applied separately for each participant. The grand average was then obtained by averaging all estimates for all participants.

Here, the main configuration parameters for the GLM are (1) the time intervals for estimated signals and (2) the time interval of the epoch for the observed signal $x\left( t \right)$, providing the number of samples⁶ (i.e.$~{N_s}$, ${N_f}$and ${N_e}$) for the evoked potential $s(t)$, $f(t)$ and observed signal $x(t)$, respectively. The estimation window of the potential $s(t)$ extended from 100 ms (baseline computation between − 100 and 0 ms) before text onset to 700 ms⁷ after. Thus, the total duration for the potential $s(t)$ was 800 ms, thereby defining the number ${N_s}$ of samples. The estimation window of the potential $fp(t)$ extended from 200 ms (baseline computation between − 200 and − 100 ms) before first fixation onset to 920 ms after, for the RM condition (840 ms after for the RD condition). This time value was set to capture the first four fixation onsets, that is, the mean IFI between the first and fourth fixations (see Table 4) was 768 ms, rounded up to 770 + 150 ms (690 + 150 ms for the RD condition). The total duration for the potential $fp(t)$ was 1120 ms for the RM condition (1050 ms for the RD condition), thereby defining the number ${N_s}$ of samples. The estimation window for the saccadic potential $s{p_c}(t)$ ($c=1..4$) was configured from 50 ms (baseline computation between − 50 and − 10 ms) before saccade onset to 200 ms after. The epoch duration for the observed signal $x(t)$ was set from 100 ms before text onset to 1150 ms after for the RM condition (1050 ms for the RD condition). These values were chosen as a function of the latency of the fourth fixation (+ 150 ms). The number ${N_e}$ of samples of the observed signals was then defined.