1. Introduction

The ability to allocate and maintain visual attention enables the flexible regulation of perception and action that is characteristic of strategic behaviour (1,2). The capacity to pay attention develops rapidly over the early years of life (3), and individual differences in early attention predict long-term cognitive and clinical outcomes (4,5). Recent new methodological advances such as naturalistic neuroimaging are allowing us to build on previous research using lab-based behavioural experiments and animal studies.

The development of attention is traditionally conceptualised as the product of interactions among different systems at different levels of maturity (1,6,7). Traditionally, the earliest subcomponent of attention to develop is thought to be the arousal/ alertness subcomponent, mediated via brainstem reticular activating systems centred on the locus coeruleus (LC) and instantiated primarily via norepinephrine neurotransmitter systems (6). In young infants, alertness is more readily initiated by exogenous events (8); over time, infants gain the ability to both attain and maintain an alert state even in absence of external stimulation. Areas around the brainstem (including the LC) are thought to be some of the earliest to become functionally mature (9,10). Consequently, the relative influence of this subcomponent of attention is thought to be strongest during early development (6).

Behaviourally, the arousal/ alertness subcomponent of attention is thought to reflect a state of anticipatory readiness, or alertness for stimulus input (6). Arousal is generally measured indirectly, via proxy measures of autonomic nervous system activity such as heart rate (11). Heart rate has been extensively studied in the context of infant attention (3,12,13). During anticipatory readiness, we know that reorientations of visual attention take place periodically, clustered around a preferred modal reorientation rate (14–18). This may reflect rhythmic activity in the central nervous system (19).

With time, it is thought that looking behaviours become increasingly modulated by higher-level executive processes that reflect the infant’s internal states, motivation, comprehension, and goals (2,20,21). Behaviourally, this increase in endogenous or internally directed attention has been shown as: a developmental increase in the degree to which attentional engagement is accompanied by decreases in distractibility (22,23); an increase in selective attention as measured indirectly, using the blink reflex (24); and differences in the trajectory of how attention durations to simple vs complex stimuli change over developmental time (25).

Other research that used experimenter-controlled, screen-based tasks to examine neural correlates of attention has examined changes in the power spectral density (PSD) of EEG oscillations, in particular infants’ theta (3–6Hz) rhythm, which increases during active, anticipatory, and exploratory behaviour (26–31). Together, these studies suggest that the expression of theta during attention-eliciting episodes could signify the engagement of neural networks related to executive attention (28,29,31). Similarly, other studies have reported decreases in alpha band activity under conditions of increased attention (32,33). Both theta and alpha effects are now widely known in the literature as “theta synchronization” and “alpha desynchronization” (29).

How children allocate their attention in experimenter-controlled, screen-based lab tasks differs, however, from actual real-world attention in several ways (34–36). For example, the real-world is interactive and manipulable, and so how we interact with the world determines what information we, in turn, receive from it: experiences generate behaviours (37). While lab-based studies can be made interactive (e.g., 38,39), how infants actively and freely initiate and self-structure their attention remains unexplored (40).

The present study aims to examine developmental changes in the relationship between autonomic arousal, cortical activity, and attention in real-world settings. To do this, we first explored how naturalistic attention patterns (measured via looking durations to play objects) from a solo play interaction change between 5 and 10 months. Then, we explored temporal relations between changes in infant arousal (measured via heart rate) and attention episodes in typical 5- and 10-month-olds infants. Finally, we investigated changes in EEG theta power relative to attention episodes, and changes in EEG theta relative to arousal (see Figure 1). As attentional systems mature and brain regions become increasingly specialised (41,42), we expected to see both a developmental increase in attention towards play objects and a developmental shift in the way different mechanisms (i.e., arousal/ alertness vs. executive attention subsystems) drive attention. To measure attention we used looking time, an approach that is known to have several limitations (12,43,44). For example, we cannot differentiate based on looking time alone whether overt and covert attention are coupled or decoupled (see 3.1).

Figure 1.

Our first set of analyses examined attentional inertia (the phenomenon that, as individuals become progressively more engaged with an object, their attention progressively increases) as a measure of internally driven attentional engagement (3,45,46). We tested whether attentional inertia influenced attentional behaviours more strongly at 10 months compared to 5 months. To do so, we calculated both the Autocorrelation Function (ACF) and the survival probability of spontaneously occurring attention episodes during play (analysis 1). The ACF allowed us to quantify the rate of change of spontaneous attention durations. A faster rate of change would indicate lower attentional inertia. The survival probability, on the other hand, allowed us to quantify the probability between looking (i.e., paying attention) and looking away. A slower decrease in the probability of an attention episode surviving would indicate increased attention engagement and decreased distractibility by other stimuli. We hypothesised that, as slow-varying fluctuations in endogenous interest or engagement start to influence looking behaviour more strongly over time, 10-month-old infants would show increased endogenous attention control indexed by a slower rate of change of attentiveness and slower decreases in the survival probability. We also predicted that we would be able to identify periodic attentional reorientations during early as well as later development (17,47); later in development, however, we predicted that infants would be more likely to extend visual fixations beyond their modal periodic reorientation rate, possibly indicating a greater or more efficient integration of attention and gaze shifting (48), and that attention duration episodes would be longer overall.

Next, in order to assess the link between lower-level mechanisms of autonomic arousal and attention, we calculated cross-correlations between autonomic arousal (indexed via heart rate) and attention episodes across the entire play session for both 5- and 10-months olds. This allowed us to examine whether arousal changes tend to forward-predict changes in attention, or vice versa (analysis 2). Based on previous research (49), we hypothesized that periods of elevated autonomic arousal would associate with, and forward-predict, shorter attention episodes. We also predicted that such relationship would weaken with time due to the maturation of cortical attentional systems.

We used a similar approach to examine developmental changes in the relationship between neural markers of executive attention and real-world attention behaviours. We were interested to examine whether neural changes (indexed by theta power) anticipate subsequent attentional behaviour shifts (48); or, whether neural processes after the attention shift relate to increases in infants’ attention engagement. To test this, we conducted three analyses. First, we analysed neural activity across a range of time windows both before and after the onsets of new attention episodes and performed linear mixed effect models to examine how neural activity before and after attention onset associated with the subsequent durations of those episodes (analysis 3). Second, we examined changes in neural activity during individual attention episodes (analysis 4). Finally, we used cross-correlations to examine whether, across the entire dataset, neural markers tend to forward-predict changes in attention, or vice versa (analysis 5). We predicted that the associations between neural markers of executive attention and real-world attention behaviours would become stronger with increasing age (i.e., theta activity would show a stronger predictive relation with infants’ attentional behaviours at 10 months, as evidence of increased modulatory power from the executive attention system on infants’ attention).

Finally, we examined whether there were any interdependencies between autonomic arousal and theta activity. We had no predictions for how this relationship would change over time.

2. Results

2.1 Analysis 1: Developmental changes in attention

Our first set of analyses examined attentional inertia as a measure of internally driven attentional engagement. Attention inertia is the phenomenon that, as individuals become progressively more engaged with an object, their attention progressively increases. We tested whether attention inertia is stronger at 10 months compared to 5 months. To do so, we calculated the ACF and the survival probability of spontaneously occurring attention episodes during play to quantify the rate of change of spontaneous attention durations and the probability between looking (i.e., paying attention) and looking away, respectively.

Initially, we conducted four descriptive analyses to test how attention and inattention durations and reorientations change over both the course of the solo play interactions and developmental time. First, we tested how many times per minute 5- and 10-months-old infants redirected their attention from one object to the other. We found that, on average, 5-month-old infants performed significantly more both attentive (t(10)= 4.346, p=0.001) and inattentive (t(10)= 4.202, p=0.002) reorientations during the solo play interaction than 10-months-old infants (Figure 2A). When we looked at how attention reorientations changed during the course of the solo play episode, we found that 5-month-old infants performed consistently more looks than 10-month-olds throughout the interaction even though the number of looks per minute decreased over the course of the interaction for both age groups (Figure 2B, and Figure S1A and B).

Figure 2.

Second, we investigated the average duration that 5- and 10-month-old infants spent in attentive and in inattentive states during the solo play interaction and minute by minute (Figure 2C and D respectively). In general, infants’ attention durations toward play objects at 10 months were longer (t(10)= -2.787, p=0.019). At 5 months, moments of inattention were longer than moments spent looking towards the object (t(10) = -3.749, p=0.003). Overall, at 10 months, but not 5 months, infants spent more time in attentive compared to inattentive states (t(58)= 10, p<0.001) (Figure 2E). We then calculated a best fit line, individual by individual, to look at how average attention duration changed within the session (see Figure S1C and D). We found no significant differences in the way attention duration changed during the interaction between the two age groups (Figure S1C and D).

Third, we explored the distribution of looks towards the objects (Figure 2F). At both ages, attention durations shorter than or equal to 5 seconds follow a positively skewed lognormal distribution, with modal attention durations in the 0.5 – 0.6 second range. Modal attention durations were significantly lower at 5 than at 10 months (t(58)= 2.211, p=0.03). Finally, the right plot of Figure 2F shows extended attention episodes. There was an increasing amount of such looks with increasing age.

Following the descriptive statistics, we calculated both the ACF and the survival probability of the looking behaviour (Figure 3). First, we used time-series analyses to examine the rate of change of attention durations, relative to itself. We calculated the ACF of the attention durations at both time points (more details in 4.3.5.2). The ACF indexes the cross-correlation of a measure with itself at different lag-intervals in time (46). ACF values were obtained from 0 to 10 seconds lag, in steps of 500 milliseconds. As shown in Figure 3A, the ACF of the time series looking behaviour fell off more sharply at 5 months than at 10 months. The ACF values were compared across ages using independent sample t-tests. From lag +500 milliseconds to 10 seconds, 10-months-old infants showed significantly higher correlation values than 5- months-old infants.

Figure 3.

Second, we performed a survival analysis by calculating the survival probability function of the looking behaviour towards the objects at both time points. The survival probability function is the probability that an attention episode survives longer than a certain time. As shown in Figure 3B, the survival probability of a look decreased abruptly at the beginning, for the very short looks, and flattened as looks got longer. The differences in the speed at which the survival probability decreased can be seen more clearly by calculating the derivative of the survival probability (Figure 3C). To compare survival among the two groups, we performed the log rank test using the Matlab function ‘Logrank’ (50). The results for the log-rank test rejected the null hypothesis (p<0.001) indicating that the survival curves for looking behaviour at 5 months and 10 months were significantly different. Notably, the likelihood of a look ending is more tightly clustered around the modal value of 0.5 seconds at 5 months.

Overall, our results showed that older infants demonstrated to have a slower-changing profile of attention with longer attention episodes overall (Figure 2A and B, Figure 3A). At both ages, there was evidence for a preferred modal reorientation rate in the 0.5–0.6 second range, which was slightly faster at 5 months than 10 months (Figure 2F). Attention durations were more tightly clustered around the modal value at 5 months. At 10 months, attention episodes were more likely to be extended beyond the preferred modal reorientation rate than at 5 months (Figure 3B and C).

2.2 Analysis 2. Auto- and cross-correlation analyses between infant autonomic arousal and attention

In this section we investigated the relationship between changes in infant autonomic arousal (indexed by heart rate activity) and their associations with moment-to-moment changes in attention (indexed as a continuum of looking durations to play objects vs. elsewhere).

Figure 4 (A and C) shows the results of the autocorrelation analyses for autonomic arousal at 5 and 10 months of age respectively. Significant autocorrelations were observed at relatively short lags around t=0 (from -4 to +4s) at both ages. Figure 4 (B and D) shows the results of the cross-correlation analysis between autonomic arousal and attention at 5 and 10 months of age respectively. The negative values indicate that, at 5 months, lower heart rate forward-predicted increased looking durations from lags between -9 to -2 seconds (i.e., lower heart rate at time t significantly associated with increased attention at time t+9 seconds). The same pattern was present but not significant at 10 months. The asymmetry of this cluster around the lag t=0 indicates that changes in heart rate tended to forward-predict changes in attention more than vice versa.

Figure 4.

2.3 Analysis 3. Calculation of neural power changes around an attention episode

We used linear mixed effects models to examine the associations between the length of each attention episode (i.e., looking duration to any of the play objects) and relative theta power at different time windows relative to the onset of that attention episode (see Figure 5). At 10 months, relative theta power in the time window of 0 to +1000msec and +1000 to +2000msec after onset of a new attention episode predicted the subsequent duration of that attention episode. At 5 months, the same relationships were not significant. We found no evidence of neural activity before the start of an attention episode forward-predicting the length of that attention episode at any time point (Figure 5).

Figure 5.

The final number of accepted trials (i.e., attention episodes) in the analyses varied across the three timewindows immediately after the onset of each look. More trials were obtained for the first window (total number of looks at 5 months was 790, and 411 at 10 months) than the second (total number of looks at 5 months was 473, and 336 at 10 months) and the third (total number of looks at 5 months was 301, and 277 at 10 months). All three conditions ended up with enough number of clean trials that was greater than the recommended number of trials in the infant EEG literature (51–53). Thus, the differences between the number of trials for each time window are not expected to contribute to the results described above. However, we repeated this analysis by matching the number of attention episodes at 5 months to the ones analysed at 12 months. We found no differences in the results (see Figure S2).

2.4 Analysis 4. Calculation of neural power changes within an attention episode

In addition to the previous analyses, which examined the associations between the length of each attention episode and relative theta power at different time windows relative to the onset of that attention episode, we also wished to examine whether power at the theta and alpha band changed significantly during an attentional look (i.e., any look at a play object) (Figure 6). Relative theta was analysed as a function of these three factors: time within an attentional episode, brain areas and age with a 3-way ANOVA (Figure 6). There was no statistically significant interaction between the three factors. However, the analysis revealed two simple two-way interactions: one between time within an attention episode and age, F(2) = 5.58, p < .005 and the other between channel cluster and age F(2) = 11.98, p < .001. Next, we performed a multiple comparison test to find out which groups of factors were significantly different. Results are shown in table S1-S3. A follow up analysis showed a significant effect of “time within an attentional episode”: 10-months- old infants had greater theta during the third-to-fourth second into the look (middle) than the first second (start) in both the central and the frontal poles. These effects were not present in 5-months-old infants. Similarly, relative alpha was also analysed as a function of these three factors: time within and attentional, brain areas and age with a 3-way ANOVA. We found no statistically significant interactions. Results are shown in Figure S3.

Figure 6.

Again, the final number of accepted trials (i.e., attention episodes) in the analyses varied across the three time-windows into each look. More trials were obtained for the first-second window (total number of looks at 5 months was 791, and 415 at 10 months) than the third-to-fourth second (total number of looks was 172 at 5 months, and 194 at 10 months) and the last second before look termination (total number of looks was 476 at 5 months, and 337 at 10 months).

2.5 Analysis 5. Auto- and cross-correlation analyses between infant theta activity and attention

In this section we investigated the relationship between dynamic changes in infant endogenous brain activity and their associations with moment-to-moment changes in attention (measured as a continuum of looking time durations to play objects vs. elsewhere). Figure 7A and D shows the results of the autocorrelation analyses for infant theta activity. Figure 7B and E shows the results for the cross-correlation analyses between infant theta activity and infant attention. Cluster-based permutation analysis revealed a significant positive association between the two variables (marked with a red line) at 10 months around time lag=0. More specifically, increases in infant theta activity at 10 months were significantly correlated with fluctuations in infant attention (Figure 7E). No associations were found between theta activity and infant attention at 5 months of age. Interpreting the exact time intervals over which a cross-correlation is significant is challenging due to the autocorrelation in the data (54,55), but the fact that the significance window is asymmetric around time 0 indicates a temporally specific relationship between infant attention and theta power, such that attention forward-predicts theta power more than vice versa.

Figure 7.

Finally, to test whether there were any interdependencies between autonomic arousal and brain activity, we performed a cross-correlation analysis between these two variables. We found a significant cluster at 10 months (Figure 7F) but not at 5 months (Figure 7C). The asymmetry of this cluster around t=0 indicated that changes in brain activity tended to precede changes in autonomic arousal more than vice versa.

3. Discussion

We examined developmental changes in the physiological and neural correlates of real-world attention patterns during early development. To do so, we measured attention durations (to an accuracy of 50Hz), along with cortical neural activity (EEG) and autonomic arousal (via ECG) from typical 5- and 10-month-old infants playing alone while seated at a tabletop with 3 toys. This age range is a key period for early cognitive development, as differential patterns of brain development (10) drive a transition from primarily subcortical to cortical control (56), and early-emerging atypicalities can have life-long consequences (57,58). However, many of the mechanisms that drive early development remain unclear.

From Analysis 1 we found that infants at both ages showed a preferred rate of reorientation (i.e. their visual attention took place periodically). The modal durations of attention episodes towards different play objects were in the 0.5–0.6 second range at both ages but were lower at 5 months (Figure 2F, 2C). This contrasts with analyses of micro-level fixation durations (time intervals between individual refoveating eye movements), which decreases from early infancy (~0.5 secs) through to later infancy (~0.4 secs) through to adulthood (~0.3 secs) (15,18). Research with adults suggests that the minimum time necessary to plan and executive a saccade is ~80msecs in adults (16). Although the equivalent figure is not known in infancy, the fact that modal attention durations towards objects were shorter at 5 months than 10 months, whereas fixation durations decrease with age, makes it likely that the figures we observed do not simply indicate that infants were reorienting at the fastest speed possible, but rather were reorienting according to a preferred modal reorientation rate (15).

The survival analysis showed that, at both ages, looks were fragile early in their existence and most likely to terminate in the <1 second range (45) but the speed at which the survival probability curve decreased was faster at 5 months, meaning that the probability of a look lasting longer than time t was lower at 5 months. Richards and colleagues have found similar relationships in infants in both lab-based and naturalistic settings (45). Overall, attention durations were shorter at 5 months; this faster-changing pattern of attention to the object was also reflected in the ACF of their looking behaviour, which decreased significantly faster, showing lower overall self-similarity. Collectively, these data fit well with what we know about the development of attention. With time, we seem to observe a higher-level control of attention that allowed older infants to prioritize the task at hand – learning about/ exploring the toys – as well as to inhibit the tendency to shift attention away from an interesting task (6,20,25). Alternatively, longer attention episodes might arise because children physically manipulate objects, bringing objects closer to themselves which makes them more exogenously salient (37,59). In this case, then the infant’s increased looking behaviour would be the result of increased exogenous attentional capture rather than an increase in endogenous attention control (46).

Analysis 2 examined how dynamic fluctuations in autonomic arousal relate to moment-to-moment changes in attention. Consistent with previous work (13,49,60), the average concurrent correlation between autonomic arousal and attention was negative at both age points, indicating that lower arousal was associated with increased likelihood of attention. Such links have been considered within the developmental attention regulation literature, where increases in arousal are thought to lead to distraction or difficulties focusing attention, and vice versa (61). We also found that arousal levels were significantly forward predictive of attention at 5 months but not at 10 months (Figure 4B and D). Theoretically, if attentional episodes drive decelerations in the heart rate (12), and older infants show longer attentional episodes on average, then one could hypothesise that older infants ought also to show a more stable pattern (i.e., higher autocorrelations) in their heart rate fluctuations than younger infants. However, this was not what we observed (Figure 4A and C). Overall, the much shorter attention durations observed in this setting, compared with screen-based TV viewing (12), means that heart rate decelerations relative to individual attention episodes were observed infrequently in our data. However, our data did suggest, consistent with previous research, that at 5 months, changes in autonomic arousal forward-predict subsequent changes in attention.

In Analysis 3 we examined the associations between attention episode durations and theta power either before, or after, onset of that attention episode. At 10 but not 5 months, increased theta during the period immediately after the onset of a new attention episode (0–2000msec) forward-predicted the subsequent length of that attention episode (Figure 5). At neither age, however, did cortical neural activity before the onset of an attention episode forward-predicted attention durations.

In Analysis 4 we examined whether cortical neural activity changed significantly during an attention episode. Consistent with previous research (29), theta power in central and frontal electrodes increased significantly during an attention episode at 10 months, but not at 5 months (Figure 6). Contrary to our expectations, we did not find a link between attenuated alpha during an attention episode at any age (see Figure S3).

In Analysis 5 we investigated the relationship between dynamic changes in infants’ theta activity and moment-to-moment changes in attention. We identified a significant positive association between infant theta activity and infant attention at 10 months but not at 5 months (Figure 7B and E). Interpreting the exact time intervals over which a cross-correlation is significant is challenging (54,55), but the asymmetry of the cluster around time 0 indicates that attention forward-predicted theta power more than vice versa, consistent with the findings from Analysis 3.

These findings are consistent with previous research suggesting that, by 10 months, but not during early infancy, theta oscillations increase during sustained attention and encoding (29,41,62) and associate with longer attentional periods (63). Importantly, though, we found no evidence that endogenous neural markers before the onset of an attention episode forward-predict the length of an attentional episode at either age. Instead, what we found suggests that neural activity shortly after the onset of an attention episode forward-predicts the length of that episode. One possible interpretation of this is that neural activity associates with the maintenance more than the initiation of attentional behaviours (64).

Finally, we examined the relationship between theta power and autonomic arousal (Figure 7C and F). A cross-correlation analysis found a negative forward-predictive relationship between the two, such that increases in theta forward-predicted decreases in autonomic arousal at 10 months, but not at 5 months. This suggests that changes in the brain activity could be modulating subcortical changes (i.e., changes in the heart rate) and may thus be able to initiate or maintain states of arousal that are common to vigilant or sustained attentional states (6,12). Overall, it appears that, by 10 months, the different substrates of attention are more inter-linked, and stronger associations are emerging between behaviour, cortical activity, and autonomic arousal (65).

In summary, our results suggest that, earlier in development, attentional episodes are more influenced by lower-order endogenous factors such as a general arousal system (3,12) - that might reflect a stronger influence of subcortical structures over the modulation of attention - and show a preferred modal reorientation timer - which characterises infants’ attention shifting more strongly. Such factors would also be present at older ages; however, their association with attention would weaken over developmental time due to the maturation in cortical attentional areas thought to take place throughout the first year of life. Later in infancy, cortical neural activity reliably changes during attention episodes, but does not forward-predict attention at either age; rather, it seems that neural changes associate with the maintenance more than the initiation of attentional behaviours. Overall, the modulation of attention seems to involve both arousalbased and cortical processes. With developmental time, however, the latter increases its control over the modulation of both (i.e., overt attentional behaviours and arousal), resulting in a more inter-linked system where associations between attentional systems are stronger. Theoretically, this is consistent with what we know about the development of executive attention from experimental and neuroanatomical studies.

3.1 Limitations and strengths

Our findings should be interpreted with consideration to a number of limitations of the study. First, our events of interest are intrinsically linked with one of the biggest EEG artefacts (i.e., eye movements), and so it is possible that residual artifact in the EEG signal may have contaminated our data. However, our data were processed using algorithms specially designed to clean naturalistic EEG data (66,67), and previous analyses suggest that the electrode locations and frequency bands that we examined should be least affected by artifact, compared with more anterior locations and higher and lower frequencies (68). Additionally, our analyses were carefully designed to preclude this potential confound. First, our analyses compare events that we know share the same level of artefact/ noise (i.e., saccades at 5 months old contribute to comparable noise levels than at 12 months old (69); second, analysis 3 and 4 are time-locked to a saccade to eliminate the possibility that saccadic frequency may have influenced our results; and third, other research (70) suggest that artifact associated with saccades disappears within ~300msecs, whereas the associations between theta and look duration lasts much longer than this, up to ~6 seconds. For all this, we consider that the possibility that our results may have been caused by infants’ saccades is unlikely.

Second, while the study of theta and alpha activity can offer insights into infants’ intrinsically guided attention beyond its behavioural manifestations (26), attributing functional significances to particular frequency bands can be risky, especially when the ways in which they are controlled and the extent to which they interact across attention-related brain networks, still remain largely unknown (71). Similarly, the fact that certain frequency bands seem to covary with attention does not exclude the possibility of their correlation with other processes as well.

Third, the use of different EEG systems (32- vs. 64-channel BioSemi gel-based ActiveTwo) and age groups might have contributed to the differences we observed over time. However, we compared the EEG signal quality between groups and found no significant differences (Table S5–6, Figure S4).

Fourth, we used a different set of toys at the two ages (see Figure S5). Consequently, this introduced a new source of variation (i.e., toy characteristics) that might have contributed to any of the observed differences (20). However, we chose to present developmentally appropriate stimuli at the two ages to ensure that the cognitive demands were similar at the two ages. Thus, while still possible, it is unlikely that the developmental differences observed in the current study might be due to differences in the amount of information processing on the part of the infant and/ or the “interestingness” of the toys.

Fifth, it is worth mentioning that, while infants gather information about their world through aggressive visual foraging, looking and attending are not synonymous. Previous research has shown that covert shifts of attention can occur without overt shifts of gaze by 4–6 months of age (72,73). However, the current study has focused exclusively on overt attention.

Finally, our laboratory setting was a novel environment for our participants and might have elicited behaviours that are different from the ones that develop at home. However, it still represents a significant advancement relative to other screen-based or highly controlled experimental tasks.

4. Materials and Methods

4.1 Experimental Design

Looking behaviour, EEG and ECG data were collected from mothers and their infants at two age points: 5 and 10 months while playing alone. At 5 months, infants were seated either in a highchair or on a researcher’s lap and a table was positioned in front so that toys on the table were within easy reach (see Figure 1). To reduce infant’s stress, mothers were present in the room but moved to another smaller table on the right side of the original table and given an identical set of toys which they played with in parallel. A wooden divider was positioned between the two tables to prevent infants from seeing the objects with which their mothers were playing. At 10 months, the same procedure was used but the divider was positioned across the midline of the table and the adult participants were seated directly opposite the infants. In both situations, mothers and infants had direct line of sight to one another but neither could see the others’ toys on the table.

The same three age-appropriate toys were always used for each age group. These were small and relatively engaging (see Figure S5). During the solo play interaction, one of the researchers sat behind the infant to collect the toys that fell on the floor (either because the infants threw them or because they fell from their hands) and brought them back on the table. Mothers were allowed to speak during the interaction but were instructed not to name the toys they were playing with to prevent infants from the influence of any exogenous parental’ influence. In the Supplementary Materials, we present a set of analyses that preclude the possibility of maternal influence on infants’ behaviour and demonstrate that the impact of the mothers on the infants’ behaviour did not differ between age groups (SI 3). The average duration of the interactions with usable EEG/ ECG data did not differ significantly between 5 and 10 months (interactions with EEG (average duration at 5m = 292.4s, and 10m = 250.1s, t(46)= -1.85, p= 0.07); interactions with ECG data (average duration at 5m = 351.2s, and 10m = 317.9s, t(40)= -1.1, p= 0.27)).

The interactions were filmed using three Canon LEGRIA HF R806 camcorders recording at 50 fps. At 5 months, one camera was placed in front of the infant and another one was placed in front of the mother. At 10 months, two cameras faced the infant: one placed on the left of the divider, and one on the right. The other camera faced the mother and was positioned just behind the right side of the divider. All cameras were placed so that the infant’s and the mother’s gaze, as well as the three toys placed on the table, were always visible.

Brain activity was recorded using a 64-channel at 5 months and a 32-channel at 10 months, BioSemi gelbased ActiveTwo system with a sampling rate of 512Hz with no online filtering using Actiview Software.

Heart rate activity was recorded using a BioPac™ (Santa Barbara, CA) system recording at 2000Hz. ECG was recorded using disposable Ag-Cl electrodes placed in a lead II position.

4.2 Participants

Participants were typically developing infants and their mothers. The catchment area for this study was East London, including boroughs such as Tower Hamlets, Hackney and Newham. Participants were recruited postnatally through advertisements at local baby groups and local preschools/ nurseries. We also operated a word-of-mouth approach, asking parents who got involved to ask if their local networks would be interested in participating. Ethical approval was obtained from the University of East London ethics committee (application ID: ETH2021-0076). Informed consent, and consent to publish, was obtained by the caregivers of the infants tested.

Initial exclusion criteria included complex medical conditions (e.g., heart rate condition, neurological/ genetic abnormality), known developmental delays, prematurity, uncorrected vision difficulties, and parents below 18 years of age. Further exclusion criteria as well as final numbers of data included in each of the analyses for both samples are summarised in Table S7. The final sample included 12 infant females and 19 infant males at 5 months and 14 infant females and 15 infant males at 10 months. Data was analysed in a cross-sectional manner. Average age was 5.32 months (std = 0.58) and 10.49 months (std = 0.87). This is the first time that any of this data has been analysed and reported.

Since the analyses are performed conducted relative to specific events (such as the frequency and duration of looks to objects), and each participant averaged around 10 looks per minute at 5 months and approximately 4 looks at 10 months (see Figure 1A and B), we believe that the relatively low N for this study is balanced by the considerable amount of data points accessible.

4.3 Data processing and Statistical Analysis

4.3.1 Synchronisation between behavioural and EEG/ ECG data

The cameras were synchronised to the EEG and ECG via radio frequency (RF) receiver LED boxes attached to each camera. The RF boxes received trigger signals from a single source (computer running Matlab) at the beginning and end of the play session, and concurrently emitted light impulses, visible from each camera. At the same time, triggers were sent and stored in the Actiview Software and recorded to the EEG data as well as to the Acknowledge Software and recorded to the ECG data.

The video coding and EEG/ ECG data synchronisation was done by aligning the times of the LED lights and the EEG/ ECG triggers. We also checked for dropped/missing frames by checking that the time between the LED lights matched the times between the EEG/ ECG triggers.

4.3.2 Video coding

The looking behaviour of the infants was manually coded offline on a frame-by-frame basis, at 50fps. The start of a look was considered to be the first frame in which the gaze was static after moving to a new location. The following categories of gaze were coded: looks to objects (where the infant was focussing on one of the three objects), looks to partner (where the infant was looking at their partner), inattentive (where the infant was not looking to any of the objects nor the partner) and uncodable. Uncodable moments included periods where: 1) the infant’s gaze was blocked or obscured by an object and/or their own hands, 2) their eyes were outside the camera frame, and/ or 3) a researcher was within the camera frame and the infant turned to them and/or realised a researcher was around. Video coding was completed by three coders, who were trained by the first author. To assess inter-rater reliability, ~15% of our data (10 datasets) were double-coded by a second coder and Cohen’s kappa was calculated. There was moderate agreement (K = 0.581, std= 0.183) (74). Due to the unusual nature of our behavioural coding (with gaze coded across many 20ms bins) the interrater reliability is heavily contingent on how we calculate it. We chose to report the most stringent calculator of inter-rater reliability.

Looking behaviour data was then processed such that any look preceding and following an “uncodable” period was NaN-ed and excluded from further analyses. Similarly, both the first and the last look of every interaction were also NaN-ed and excluded from further analyses.

4.3.3 EEG artefact rejection and pre-processing

EEG data was pre-processed and cleaned from oculomotor and other contaminatory artefacts using a fully automatic artefact rejection procedure specially designed for naturalistic infant EEG data by Mariott Haresign (66), building on previous related work (59, 60). Briefly, this involved the following steps: 1) data were high-pass filtered at 1Hz, 2) line noise at 50Hz was eliminated using the EEGLAB function clean_line.m, 3) data were low-pass filtered at 20Hz, 4) the data were referenced to a robust average reference 5) noisy channels were rejected using the EEGLAB function pop_rejchan.m, 6) the channels identified in the previous stage were then interpolated back, using the EEGLAB function eeg_interp.m, 7) continuous data were automatically rejected (NaN-ed) in a sliding 1s epoch based on a percentage of bad channels (set here at 70% of channels) that exceed 5 standard deviations of the mean channel EEG power, and 8) Independent Component Analyses (ICA) were computed on the continuous data using the EEGLAB function runica.m. Only participants with fewer than 30% of channels interpolated at 5 months and 25% at 10 months (step 6) made it to the final step (step 8, ICA) and final analyses. To compare the quality of the EEG data at 5 and 10 months we performed a series of analyses on percentage of channels interpolated, total segments removed (i.e., zeroed out) and total ICA components rejected (see Table S5–6 and Figure S4).

The higher density array was down sampled so that all the EEG analyses described below used only shared channels between the 32- and the 64- channel EEG systems. We selected three main clusters of electrodes for our analyses: Frontal channels (‘Fp1’, ‘Fp2’, ‘AF3’, ‘AF4’, ‘Fz’), Central channels (‘FC1’, ‘FC2’, ‘C3’, ‘Cz’, ‘C4’, ‘CP1’, ‘CP2’), and Occipital channels (‘PO3’, PO4’, O1’, ‘Oz’, ‘O1’) (see Figure S6).

4.3.4 Heart Rate - Beats per minute

R-peak identification was performed using the in-built Matlab function ‘findpeaks’. The minimum peak height was manually defined as a simple amplitude threshold after visualising the raw data, minimum peak distance, instead, was set at 230msec. Following this, automatic artefact rejection was performed by excluding those beats showing an inter-beat interval (IBI) <330 or >750msec (i.e., allowing a minimum of ~80BPM and maximum of ~180BPM), and by excluding those samples showing a rate of change of IBI greater than 90msec between samples. Next, we converted IBI values into beats-per-minute (BPM) values and removed outliers in the BPM time series. These were defined as values falling 2.5 interquartile ranges above the upper quartile and below the lower quartile. Outliers were then interpolated using the Matlab inbuilt function ‘fillmissing’ with the ‘spline’ method. Finally, we epoched the data into one-second epochs by averaging all the BPM values comprised in each one-second epoch.

4.3.5 Analysis 1. Developmental changes in attention

4.3.5.1 Overt attention and inattention extraction

The aim of our analysis was to identify moments where the infants paid attention to (i.e. looked at) any of the play objects as opposed to inattentive moments. Accordingly, all looks to object and inattentive looks were selected and categorised as attentional and inattentive episodes, respectively. Looks to partner were excluded from all analyses.

Following this, we extracted the first and last frame of all looks of interest (i.e., looks to objects and inattentive looks). To calculate attention and inattention durations, we subtracted the last frame from the first frame of each look of interest and divided it by the sampling rate (i.e., 50) to convert “duration in frames” to “duration in seconds”. Attentive and inattentive reorientations (count) were calculated by counting the occurrence of each of these two attentional states and dividing it by the length of the session.

4.3.5.2 ACF of the attention duration

Here, we extracted the duration (in seconds) of all the attentional episodes that happened within the play session and zero-ed out all the non-attentional episodes. This allowed us to create a time series string with the durations of each consecutive attention episode. We then calculate the autocorrelation of that signal and repeated these steps for each behavioural dataset. Finally, we averaged the ACF values within each age group to obtain the ACF values reported in Figure 3A.

4.3.6 Analysis 2. Auto- and cross-correlation analyses between infant autonomic arousal and attention

4.3.6.1 Attention, one-second epochs

To calculate the “attention” variable, we epoched the gaze data into one-second epochs and calculated the duration (in msec) of each attention episode relative to the one-second epochs. Most epochs were coded as either 1000 (epochs where the child was attending throughout) or 0 (epochs where the child was inattentive throughout). If an attention episode started halfway through a one-second epoch, then it was coded as 500. The other non-attention episodes (i.e., inattentive looks, looks to partner) were zeroed. See Figure 1 for a schematic view of the procedure we followed to parse the looking behaviour into one-second epochs.

4.3.6.2 Cross-correlation analyses

To investigate the relationship between autonomic arousal and fluctuations in attention we performed a cross-correlation analysis between the two variables. Importantly, these analyses are not time locked to specific moments (i.e., start of an attentional event) and are conducted on two time series (i.e., attention and heart rate fluctuations) as a whole. Because of this, the strength of the overall correlation is weakened by the fact that periods of expected stronger correlation are balanced by weaker correlations where we would not necessarily expect any correlation at all.

Additionally, we also computed the autocorrelation for autonomic arousal to assess how well it predicts itself over time and evaluate its stability. All analyses were computed at lags from -30 to +30s in 1s intervals. The cross-correlations values at each time-lag were computed individually and then averaged across all participants. The procedure was identical for the autocorrelation, except that instead of examining the relationship of two different time series at variable time intervals, we assessed the relationship of one time series to itself at variable time intervals.

To assess significance of the cross-correlations, we first used bootstrapping to generate confidence intervals, using an approach that controls for the level of autocorrelation in the data. To do this, the time series data of one participant (e.g., attention of participant 1) was randomly paired with the time series data of another participant (e.g., autonomic arousal of participant 13). If the time series datasets had different lengths (due to different participants having different session lengths), we appended zeros to the end of the shorter vector to match the length of the longest vector. We then computed the cross-correlation between all the unique combinations that could be found within each sample (e.g., in a sample N=23, the maximum of unique combinations is 529). Next, the cross-correlation results in the permuted data were randomly grouped in samples that were the same size as the original data (e.g., N= 23) and averaged together. This procedure was repeated 1000 times and used to generate the 95% confidence intervals. In this way, we identified whether the observed cross-correlation values at each time interval differed significantly from chance.

Next, to control for multiple comparisons across time intervals, we used a cluster-based permutation approach. On each iteration, one permutation was compared with the 999 other permutations, and significant time-points were identified as values falling above the 97.5th centile and below the 2.5th centile (corresponding to a significance level of 0.05). We then identified the two largest clusters of significance that occurred by chance: one for positive correlation values and the other for negative correlation values. We repeated this 1000 times. Following this, we created a distribution of cluster sizes for positive and negative correlation values and took the size value corresponding to the 95th percentile in each distribution to define our cluster-size threshold. Finally, we compared the cluster sizes obtained in the observed data against the cluster-size threshold and only considered significant the ones that exceeded such threshold.

Calculating the significance levels of the autocorrelation was more straightforward. This was done by first calculating the autocorrelations based on individual datasets, and then averaging the significance values of the Spearman’s correlations at each time interval.

4.3.7 Analysis 3. Calculation of neural power changes around an attention episode

We examined the associations between the duration of infant attention episodes and infant theta changes around these looks using linear-mixed effects models. Infant attention episodes and attention duration were calculated as explained above in section 4.3.5.1.

To conduct these analyses, each infant attention episode onset (i.e., gaze shift to a different toy) was identified in the EEG signal by calculating the time from the start of the interaction (first LED) to the onset of the look (in the behavioural data) and adding it to the first EEG trigger. For each look, we extracted theta (3–6Hz) power for two time-windows immediately prior to the onset of each look (-2000 to -1000msec and -1000 to 0 msec pre-look onset) and three time-windows immediately after the onset of each look (0 to 1000ms, 1000ms to 2000msecand 2000 to 3000msec post look onset) (see Figure 1). To calculate the EEG power spectra, we use the ‘mtmfft’ method from the ft_freqanalysis function in FieldTrip, an open-source Matlab toolbox (75); http://fieldtriptoolbox.org). Extreme power values that were 4 times greater than the interquartile range were treated as outliers and excluded from further analyses (similar to 29). More detail on the amount of data available (i.e., average duration of the session per participant and number and duration of attentional episodes per minute) can be accessed in Table S1.

For each epoch, we only selected power within our cluster of central channels (similar to 60). Power at each bin was expressed as relative power, defined as the total power at a specific frequency band (e.g. 3 to 6Hz for theta) divided by the total power across all frequency bands (1 to 20Hz) during that epoch. After extracting the relative power in the theta band, we calculated separate linear mixed effects models for each of the five windows to examine the relationship between EEG power within that time window and attention duration.

4.3.8 Analysis 4. Calculation of neural power changes within an attention episode

In addition, we also wanted to look at power changes within attention episodes. Infant attention episodes and attention duration were calculated as explained above (section 4.3.5.1) and each infant look onset towards an object was identified in the EEG signal as described in analysis 3 (section 4.3.7). For each look, we extracted the first (0 to +1000msec, “start”) and third-to-fourth (3000 to 4000msec, “middle”) second into the look, and the last second (-1000msec prior to look offset to look offset, “termination”) before look termination (see Figure 1). Looks that did not make it to the full second segment were excluded from further analysis. Similarly, only looks that were longer than 5 seconds were included to the “middle” group. This was done to avoid an overlap between the activity from the “middle” and the “termination” groups.

Frequency analysis was conducted to assess the power spectral density for both theta (3–6Hz) and alpha (69Hz) frequency rhythms for each of the three time-segments. These analyses were calculated for the three prespecified clusters of channels: Frontal, Central and Occipital (see Figure S6). Again, power at each time segment was expressed as relative power.

The selection of both theta (3–6Hz) and alpha (6–9Hz) frequency bands was led by previous work using this same approach (e.g., 28, 29, 30, 37).

4.3.9 Analysis 5. Auto- and cross-correlations analyses between infant theta activity and attention

4.3.9.1 EEG relative power, one-second epochs

For this analysis, we parsed the EEG data into one-second segments and calculated the relative theta power for each one-second segment as described above (see 4.3.7).

4.3.9.2 Cross-correlation analysis

To explore whether modulations in endogenous theta activity related to fluctuations in infants’ attention, we conducted a cross-correlation analysis between infants’ relative theta and attention. Attention was calculated as described in analysis 2 (4.3.6.1, and Figure 1). Additionally, we also computed the autocorrelation for relative theta to assess how theta predicts itself over time. Again, all analyses were computed at lags from -30 to +30s in 1s intervals. Significance was assessed following the steps described in analysis 2 (4.3.6.2).

Finally, to explore interdependencies between autonomic arousal and theta activity we conducted a crosscorrelation analysis between infants’ autonomic activity and relative theta.

Acknowledgements

We want to send a special thank you to María José Peñaherrera, Mükrime Gök, Stefanie Pow, Desirèe Cardile, Georgina Harris and Tecla Zanin for their unlimited patience and dedication with data coding. We also want to thank the members of the UEL BabyDev Lab for comments and discussions on earlier drafts of this manuscript, and to all participating infants and caregivers that took part in our study.

Funding

This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. [853251 - ONACSA]) (awarded to SVW), and the Leverhulme Trust, grant RPG-2018-281 (awarded to SVW).

Ethics

Human subjects: Ethical approval was obtained from the University of East London ethics committee (application ID: ETH2021-0076). Informed consent, and consent to publish, was obtained by the caregivers of the infants tested. Specific informed consent was also obtained from the caregivers of the infants featured in Figure 1.

Author contributions

Conceptualization: SVW, EJHJ, MPA.

Data curation: IMH, MW, MPA.

Formal analysis: MPA.

Funding acquisition: SVW.

Investigation: EG, JI, PL, IMH, EAMP, NKV, MW, MPA.

Methodology: IMH, EAMP, SVW, EJHJ, MPA.

Project administration: EG, JI, PL, IMH, TJN, EAMP, NKV, MW, SVW, MPA.

Supervision: SVW, EJHJ.

Visualisation: MPA.

Writing – original draft: MPA.

Writing – review & editing: EG, JI, PL, IMH, TJN, EAMP, NKV, MW, EJHJ, SVW, MPA.

Competing interests

Authors declare that they have no competing interests.

Data and materials availability

Partial restrictions to the data and/or materials apply. Due to the personally identifiable nature of this data (video recordings from infants) the raw data will not be made publicly accessible. Researchers who wish to access the raw data should email the lead author. Permission to access the raw data will be granted as long as the applicant can guarantee that certain privacy guidelines (e.g. storing the data only on secure, encrypted servers, and a guarantee not to share it with anyone else) can be provided.