Introduction

Across developed countries 9.5 per 1,000 births are Very Preterm (VP), i.e. born between 28 and <32 gestation weeks [1]. Although survival and health prospects of VP children have improved since the 1990s [2, 3], they are still at increased risk for cognitive and learning difficulties [4–7], problem behaviour [8], and developmental disorders such as Attention Deficit with Hyperactivity Disorder [9–13].

Longitudinal studies of VP infants suggest that problems in learning and behaviour stem from difficulties in the way they regulate and control the acquisition of information from an early age [14–18]. A crucial role is played by VP infants’ anomalies in attention control, the ability to voluntarily determine what to pay attention to, and what to ignore. This ability starts to emerge around the first year of life in typically-developing infants [19, 20], and marks a transition from infants’ attention being highly reactive and directed externally by the environment, to their attention being directed voluntarily in accordance with the infants’ goals [14]. Attention control is considered a key precursor of general top-down abilities to control cognitive processes in the service of goals such as Executive Functions (EFs) [21, 22]. Thus, early appearing anomalies in attention control may cause a cascade of consequences that affect the development of learning and self-regulation [18, 23–26].

Recent studies indicate that cognitive training interventions improve key EFs skills of children who were born VP [27–30]. Applying cognitive training interventions at an earlier age may produce larger transfer of effects [31], particularly if targeted at a key skill such as attention control. A similar intervention, the Attention Control Training (ACT), has been developed in typically developing infants [32–35]. The intervention involves showing interactive cartoons on a computer screen that move and change in response to infants’ direction of gaze, thanks to an eye-tracking device that feeds information to the computer in real time. Infants engaged in the games receive “reward” presentations (e.g. display of characters making funny noises) if their visual behaviour meets increasing demands (e.g. maintaining attention on one character on the screen whilst ignoring an increasing number of visual distracters). Repeating these tasks over time provides the opportunity to practise infants’ attention control abilities in age-appropriate and engaging games.

We conducted a study to test the feasibility of delivering the ACT to VP infants [36]: in a previous study [37, 38] we presented the main results concerning feasibility (e.g. recruitment and retention): these results indicated that infants engaged in the training and assessments. Conversely, in this manuscript we investigate the direction of treatment effects across a series of infants’ outcomes using Bayesian regression. Bayesian analyses allow to update assumptions about parameters (e.g. differences between two groups) by incorporating new information [39]. Bayesian approaches differ radically from conventional ones in the way they conceptualise parameters [40, 41]. In the conventional statistical approach, dubbed the frequentist approach, parameters are considered unknown but fixed: consequently, we cannot investigate the probability of a parameter of interest. Instead, we can only investigate the probability of a parameter being equal to the observed value (or more extreme ones) in a hypothetical counter-factual scenario that assumes a pre-defined value for the parameter, as in null-hypothesis testing. Furthermore, because the distributions of parameters are unknown, validity of inference in the conventional approach relies on the assumption that the sampling distribution of the parameter of interest approximates a known distribution (e.g. a normal distribution). This assumption is often plausible only when studies collect a large amount of information, e.g. by recruiting large numbers of participants.

In the Bayesian approach parameters are assumed to be uncertain, and thus can be described by probability distributions: these distributions can, in turn, be updated by incorporating newly observed data, thus providing more credible estimates. Inference in Bayesian analyses is thus informative and valid regardless of the sample size. Furthermore, by obtaining an updated probability distribution of parameters, it is possible to report statistics like Uncertainty Intervals (UI) that intuitively represent the range of values that the parameter of interest can credibly assume. Bayesian approaches are therefore well suited to provide valid and unambiguous inference even when few data are collected, and also provide straightforward methods for estimating the probability of replicating studies’ results and power [42]: Bayesian approaches are thus apposite for the analyses of preliminary and exploratory studies or studies of small clinical samples. We started the analyses with a sceptical assumption of no differences in outcomes between treated and controls: these assumptions were updated considering the data we collected.

Methods

This study was reviewed and approved by the Health and Social Care Research Ethics Committee A (HSC REC A), Office for Research Ethics Committees Northern Ireland (ORECNI), on 09 March 2018, REC Reference: 18/NI/0010; IRAS Project ID: 237537. The study has therefore been performed in accordance with the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments, as well as national laws concerning data protection. Caregivers of infants enrolled in the study provided written informed consent to take part in the study before initiating the study. The custodian of individuals pictured in Supplementary Material Section 3 (pages 6 and 7) in S1 File has provided written informed consent (as outlined in PLOS consent form) to publish their image alongside the manuscript.

We used Bayesian regression analyses to investigate the effects of the ACT programme in a small sample (n = 10) of VP infants aged around 12 months from conception (in order to control for prematurity). Infants were randomly allocated to receiving the treatment or an active control procedure. Our overarching aim was to provide initial estimates of the intervention effects in order to inform future studies. Outcomes collected included visual attention and visual memory assessed using screen-based computerised tasks. We also investigated focused and social attention, communication skills, and infants’ temperament using naturalistic structured tasks to test transfer of effects in contexts and domains beyond those targeted by the training.

Participants

Between 2018 and 2019 we recruited infants born very preterm who were under 12 months of age (corrected age) at time of recruiting. The key exclusion criteria were presence of significant cognitive or sensory and motor difficulties, or congenital anomalies (e.g. Cerebral Palsy). In Fig 1, we report a diagram detailing numbers assessed for eligibility, participants randomised and included in the analyses: Infants were allocated with 1:1 ratio to either the Attention Control Training (ACT) or a control procedure. Overall, 12 infants were recruited, but 10 (evenly split between treatment and control) completed the training/control procedure and the post-test assessments: One infant who had been randomised to the training dropped out before the scheduled pre-test; Another infant, randomised to the control procedure, dropped out after completing the pre-test. Participants’ characteristics are reported in Table 1. We highlight that the main caregivers indicated higher than average educational attainments. Due to events beyond our control (e.g. families’ commitments), infants in the control group were older at time of testing compared to treated ones. Thus, in data analyses we controlled for infant’s age.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. CONSORT diagram.

https://doi.org/10.1371/journal.pone.0273767.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Participants’ characteristics by group allocation (n = 10).

https://doi.org/10.1371/journal.pone.0273767.t001

Results

We report descriptive results for the screen-based tasks in Table 2, and for the naturalistic tasks in Tables 3 and 4. Results concerning comparison and selection of models are presented in Table 5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Observed results of the screen-based tasks.

https://doi.org/10.1371/journal.pone.0273767.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Observed results of naturalistic attention tasks.

https://doi.org/10.1371/journal.pone.0273767.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Observed results of ESCS tasks and the Lab-Tab “Toy in a Box” task.

https://doi.org/10.1371/journal.pone.0273767.t004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Results of model comparisons using Bayesian regression.

https://doi.org/10.1371/journal.pone.0273767.t005

Screen-based visual attention tasks

Results (see Table 2) indicated that the treated improved their performance in all these tasks, save for the looks to habituation outcome. Treatment-related changes were noticeable in Visual Memory and Information Density tasks: in contrast to treated infants, the controls displayed worsened performance between pre- and post-test. Model comparisons (see Table 5) indicated that age moderated the effect of treatment for peak looks, visual memory, and attention capture: differences between treated and controls diverged with increasing age, see Fig 2. We report estimates of d parameters and scores in Table 6 and Fig 3. Overall, the results indicated negligible treatment effects in the Sustained Attention task and in the Gap-Overlap task ( = 0.02 and = - 0.02, respectively). Moderate effects favouring the treated were observed in the Visual Memory ( = 0.59) and the Information Density task, ( = 0.82): although the UIs of these estimates were large, the probability of scores in favour of the treated was 85% and 91% respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Fitted (i.e.: Estimated) d scores by treatment and age in models that suggested relevant interactions treatment by age.

Lines represent estimated scores based on the posterior distribution. The shaded areas represent 89% uncertainty intervals of the scores estimated from the posterior distrition. Dots represent the observe treated and control scores by age.

https://doi.org/10.1371/journal.pone.0273767.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Estimated distributions of scores (i.e. differences in d scores between treated and controls) by screen-based outcomes.

The dotted lines highlight a = 0, i.e. no difference in performance between treated and controls. Scores on the right side of the dotted lines indicate improved post-test perfromance of the treated compared to the controls. NOTE: These scores were estimated by sampling n = 10,000 observations from the posterior distribution of parameters estimated in each selected model (see Table 5). Where the effect of the treatment was moderated by age, scores for treated and controls were estimated by considering scores at age 13 months.

https://doi.org/10.1371/journal.pone.0273767.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Parameters estimated from the posterior distributions of models by screen-based outcomes.

https://doi.org/10.1371/journal.pone.0273767.t006

Naturalistic attention tasks

Results indicated that the trained infants displayed better performance compared to controls (see Table 3). Model comparisons indicated Age by Treatment effects for intensity of facial interest in the Lab-TAB orienting task and looks to objects during free play, see Table 5: Differences between treated and controls in facial interest diminished with age, while they increased with age when considering looks to objects, see Fig 2. Analyses indicated a moderate treatment gain ( = 0.57) in outcome intensity of facial interest, see Table 7 and Fig 4. A moderate difference in favour of the treated was also observed in infants’ looks to objects during free play, = 0.65. Although the UIs were large and imprecise, the probability of these differences favouring the treated were 82% and 84% respectively, see Fig 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Estimated distributions of scores (i.e. differences in d scores between treated and controls) by naturalsitic tasks outcomes.

The dotted lines highlight a = 0, i.e. no difference in performance between treated and controls. Scores on the right side of the dotted lines indicate improved post-test perfromance of the treated compared to the controls. NOTE: These scores were estimated by sampling n = 10,000 observations from the posterior distribution of parameters estimated in each selected model (see Table 5). Where the effect of the treatment was moderated by age, scores for treated and controls were estimated by considering scores at age 13 months.

https://doi.org/10.1371/journal.pone.0273767.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Parameters estimated from the posterior distributions of models by naturalistic tasks outcomes.

https://doi.org/10.1371/journal.pone.0273767.t007

Discussion

The overarching aim of the study presented here was to provide initial estimates of VP infants’ short-term changes across attention and socio-cognitive tasks after an early computerised cognitive training intervention. The intervention, delivered using a RCT design as part of a feasibility study, specifically targeted infants’ attention control, which is considered a foundational skill for the development of top-down abilities in the service of learning and behaviour regulation [22, 49, 50]. VP infants were approximately 12 months (corrected age) when taking part in the study, an age when attention control abilities are considered to be emerging. We hypothesised the intervention could improve infants’ performance in the short term, particularly in attention tasks more akin to the abilities trained (e.g. sustained attention). To provide a robust test of our hypotheses, in our analyses we initially assumed a sceptical view, postulating no differences between treated and controls: using Bayesian methods, these assumptions were updated considering the data we collected.

VP infants who received the training did not display noticeable gains in tasks such as habituation and disengagement. This contrasted with results from two previous studies of typically-developing infants who received the ACT programme for an equivalent length of time and number of sessions as in this study [31, 32]. In particular, while the effect of the ACT on sustained attention in our study was negligible ( = 0.02, 89% UI -1.00 to 1.05), Wass and colleagues [31] and Ballieux and colleagues [32] had reported moderate treatment effects in this task: Cohen’s d = 0.65 and Cohen’s d = 0.69, respectively. In our study we reported a negative but negligeable effect of the intervention on disengagement latencies in the Gap-Overlap task ( = - 0.02, 89% UI -1.02 to 0.98), while previous studies reported a moderate positive (Cohen’s d = 0.68, see [31]) and small positive effect (Cohen’s d = 0.35, see [32]). Future studies comparing term and VP infants’ performance in these tasks across time and different phases of the intervention are necessary to investigate attention differences and their underlying mechanisms across these groups.

Our study, however, did indicate moderate treatment effects in the ability to memorise complex visual stimuli ( = 0.59; 89% UI -0.33 to 1.52), and the ability to maintain attention on less salient and less automatically attention-eliciting stimuli presentations ( = 0.82; 89% UI -0.15 to 1.77). Moderate treatment effects were also observed in naturalistic tasks (e.g. focused attention during object play: = 0.65; 89% UI -0.43 to 1.71) and in parental reports of infants’ effortful control ( = 0.53; 89% UI -0.63 to 1.63). An aversive treatment effect ( = - 0.58; 89% UI -1.52 to 0.37) was reported insofar treated infants were more prone to protesting in a naturalistic frustrating situation: However, infants displayed low levels of protest during this task, as demonstrated by the average ratings of this behaviour being close to 0 (see Table 4).

This is the first study that trained VP infants’ attention control. Previous analyses indicated it is possible to recruit and retain VP infants and engage them in the training [37, 38]. Conversely, the analyses of infants’ performance in different tasks presented in this paper are important in providing initial estimates concerning the potential for improving early key cognitive abilities of this at-risk group. The treatment effect in visual memory was estimated to be as large as 0.59 SD units: this suggests treated infants displayed better abilities to maintain task-relevant information for brief periods of time. Performance in memory tasks in infancy can predict working memory in adolescence [51]: improvements in memory may thus be an important foundation for later attainments. A moderate treatment effect of 0.82 SD units was also estimated in infant’s abilities to overrun a preference for ‘automatically’ attention-eliciting stimuli presentations in favour of less salient ones. This ability may indicate a shift from attention being exogenously controlled to attention being controlled endogenously (from the inside) and voluntarily. The results indicate that treatment effects were moderated by infants’ age in the study (see Fig 2): improvements were larger with increasing age of the treated. Whilst early intervention may favour VP infants’ development, these results suggest the need to consider the presence of sensitive periods: VP infants may be more receptive to attention control training with increased maturation.

Moderate treatment effects observed in naturalistic tasks are suggestive of potential for transfer of effects. Treated infants displayed improvement in focused attention during naturalistic object play. Parents of infants in the intervention also reported improvements in infants’ effortful control in a validated questionnaire, a dimension that taps into infants’ focused attention (e.g. prolonged play with objects), and inhibitory control (e.g. infants’ soothing in response to parental talking). Parents were not told whether their child was receiving the treatment or the active control procedure, and the results we reported in a previous paper [37, 38] showed that parents did not accurately guess in which group their child had been allocated: we thus argue it is unlikely that parental ratings were biased by their awareness of group allocation. Overall, these findings suggest treated infants being able to focus attention during their interactions with the environment. A study with a similar population reported that preterm infants’ focused attention predicted cognitive performance in childhood [17]: replication and extension of our results would thus suggest encouraging prospects for treated infants.

A moderate treatment effect was also observed in infants’ Initiating Joint Attention (IJA), the ability to employ their direction of gaze and gestures to spontaneously share experiences with others. IJA has been linked to voluntary attention process underpinned by frontal systems [52–54]. Joint attention, in turn, is pivotal in enabling infants’ social learning [55]: further evidence of effect transfer into joint attention skills would thus strengthen the case for early training of attention control playing a gating role in the onset of further learning skills.

The results concerning treated infants’ increased protest in a frustrating situation may induce caution. However, descriptive results indicate the protest behaviour in the task was of low intensity overall. Furthermore, other studies suggest that protest in frustrating situations is a normative reaction when infants develop abilities to explore the environment [56]: infants’ protest might rather be interpreted as a sign of infants’ increased engagement with the task.

Conclusions

While the results of this study should not be over-estimated, the use of a Bayesian approach is useful in providing credible estimates of effects generated by combining the evidence collected with explicit (i.e. formalised) sceptical assumptions (see SM–Section 5 in S1 File). This approach is valuable in estimating potential effects in a small sample like ours: advances in computing systems, as well as the development of statistical packages and tutorials on the implementation of Bayesian approaches, make these more accessible for exploratory studies, studies of small clinical groups, and so on. While the traditional frequentist approach provides “impoverished” estimates of parameter values with ill-defined confidence intervals [42], Bayesian approaches provide unambiguous information about the relative credibility of parameter estimates as well as their trade-offs. The use of Bayesian approaches is also consistent with a move away from reliance on traditional null hypothesis testing which, coupled with publication bias, may have led to inflation of published effect sizes and the “reproducibility crisis” in disciplines like Psychology [59].

The results presented in this manuscript suggest moderate intervention effects in key attention and social attention domains as well as moderate probabilities of training effects being positive (see Tables 6 and 7, and Figs 3 and 4). These short-term effects were observed even though infants took part in just three training sessions for a cumulative average duration of less than 90 min of training. Overall, these results warrant further study. There is increasing evidence that at-risk populations like VP infants are responsive to cognitive training [29], and that this type of training may be particularly effective when delivered at early ages [31], indicating cognitive control training as a promising avenue of intervening in at-risk or vulnerable populations. Furthermore, while VP infants may be at risk of learning and cognitive problems due to early birth and vulnerability to brain connectivity anomalies [60, 61], biological models suggest that early experiences may modulate individuals’ reactivity to the environment [62]: VP infants’ may be more susceptible to environmental inputs and thus display increased gains from positive environmental influences [63]. Future trials will be pivotal in further investigating the effects of early cognitive interventions, as well as the durability and scope of these effects across larger and more diverse samples of VP infants.

Supporting information

Acknowledgments

We are very grateful to all the parents and their children who gave their time to take part in this study. Dr Nita Saxeena (HSC South-East Trust) and Dr Sanjeev Bali (HSC North Trust) gave a significant and generous contribution to the recruitment of participants in the study. Dr Nicola Doherty (HSC Western Trust) and Danielle Barnes contributed with insights and feedback during Steering Group meetings. We also wish to thank the TinyLife workers that enthusiastically helped with recruitment and the running of the study. We thank Aaron Patterson for his contribution to data collection, and Dr Ahmet Butun for his contribution to data coding.