Characterization of language abilities and semantic networks in very preterm children at school-age

doi:10.21203/rs.3.rs-4130846/v2

Download PDF

Article

Characterization of language abilities and semantic networks in very preterm children at school-age

https://doi.org/10.21203/rs.3.rs-4130846/v2

This work is licensed under a CC BY 4.0 License

Version 2

posted

You are reading this latest preprint version

It has been widely assessed that very preterm children (<32 weeks gestational age) present language and memory impairments compared to full-term children. However, differences in their underlying semantic memory structure have not been studied yet. Nevertheless, the way concepts are learned and organized across development relates to children’s capacities in retrieving and using information later. Therefore, the semantic memory organization could underlie several cognitive deficits existing in very preterm children. Computational mathematical models offer the possibility to characterize semantic networks through three coefficients calculated on spoken language: average shortest path length (i.e., distance between concepts), clustering (i.e., local interconnectivity), and modularity (i.e., vocabulary enrichment). Here we assessed these coefficients in 38 very preterm schoolchildren (aged 8-10 years) compared to 38 full-term schoolchildren (aged 7-10 years) based on a verbal fluency task. Using semantic network analysis, very preterm children showed a longer distance between concepts and a lower interconnectivity at a local level than full-term children. However, we found no differences between very preterm and full-term children regarding their modularity at a global level. These findings provide preliminary evidence that very preterm children demonstrate subtle impairments in the organization of their semantic network, encouraging the adaptation of the support and education they receive.

Biological sciences/Neuroscience/Cognitive neuroscience

Biological sciences/Neuroscience/Learning and memory

Biological sciences/Psychology/Human behaviour

very preterm birth

semantic network

language

Very preterm births (i.e., before 32 gestational weeks) are known to cause various adverse neurodevelopmental outcomes [1,2]. Between 3-5 years, very preterm children present more difficulties in multiple cognitive domains compared to full-term children [3,4]. Among these domains, language is often reported as a problematic [5], with long-lasting impairments commonly reported in preterm children (for reviews see for example [6] or [7]). With the exception of those without severe neonatal medical complications [8], language impairments seem to affect many subdomains. While early childhood delays are caught up in low-level language abilities (i.e., phonological processing and sight word reading) by early adolescence, high-level abilities (i.e., syntax, semantics, phonemic decoding, and verbal language memory) remain more often impaired in very preterm adolescents [9,10]. Consequently, they need about eight times more academic support in writing and reading compared to their full-term peers [5]. One aspect of these long-lasting impairments concerns their vocabulary. Where Kunnari et al. [11] found no differences in expressive vocabulary for children aged 2, Luu et al. [5] showed poorer receptive and expressive vocabulary in preterm children compared to full-term controls aged 8 and 12. However, later at the age of 16, they found that the deficit in receptive vocabulary was no longer present [9]. Moreover, while 3-year-old very preterm children uttered the same number of questions and negations, and had comparable lexical variability (i.e., number of nouns, verbs, adjectives, prepositions) compared to full -term children, they produced less varied sentence structure with, in addition, shorter and fewer words as well as fewer words roots [12]. On the other hand, preterm children also exhibit impaired reading skills and comprehension as well as delayed grammatical development (e.g., [13]). These skills rely in particular on the word comprehension [6]. Altogether these quantitative measures show that preterm children have poorer receptive and expressive language, semantics, and receptive grammar, reflecting a lower overall comprehension capacity (e.g., sentences, words, structure, verbal information) [14].

Moreover, most studies report that these impairments also affect verbal memory. At 3.5 years old, preterm children already present poorer phonological working memory (i.e., memory of words and non-words) than full-term children. Omizzolo et al. [15] discovered that despite preterm children having a good long-delay verbal recall, their short-delay verbal recall remains impaired at 7 years compared to full-term controls. This persists during adolescence, with 9-16-year-old preterm children showing weaker linguistic working memory, based on sentence recall and the establishment of semantic relationships than full-term of the same age [16].

Whereas language abilities and verbal memory have been well studied in very preterm children, the underlying complexity of semantic organization has not been explored yet. The way concepts are learned and progressively organized in memory predicts various higher-order abilities [17]. Consequently, language difficulties in very preterm schoolchildren may mirror the underlying semantic organization. Indeed, children reinforce and expand their knowledge in their early years. During their first 7 years, they acquire on average 800 root words per year and then learn 1,000 new words per year until the age of 12 [18]. Vocabulary and concepts are stored in the semantic memory network [19]. Alongside this acquisition, children also weave links between these words and gain a deeper understanding and meaning of concepts, allowing them to use abstract and conditional reasoning [20]. While gaining knowledge and expertise in a particular domain, the semantic structure becomes more cohesive and associative, thus they have better access to specific knowledge [17]. Semantic memory structure is considered to be built and influenced by experience [21], but there is a gap in knowledge concerning children who have early developmental adversity.

Recently, some studies have begun to focus on semantics in the form of networks using methods based on mathematical models and computational research (e.g., [22]). Concepts are represented by nodes (i.e., basic units of the network) which are interconnected (i.e., represented by the edges) [23]. These associations are based on different similarities, creating a network. This latter is structured by three mechanisms, the modularity (i.e., Q, subnetworks creation), the average shortest path length (i.e., ASPL, moving from one node to others), and the clustering (i.e., CC, closeness, or isolation of the nodes). These parameters influence the way we perceive and retrieve concepts. For example, the more distant the concepts are, the more difficult it is to make associations between them [22].

By investigating semantic networks in very preterm children, we aimed to examine if their verbal difficulties could relate to a problem with the organization of the semantic memory and the encoding and storage of concepts. One common and simple way to investigate semantic networks is the verbal fluency task where participants are asked to give as many words as possible in a specified category (e.g., animal, food, words beginning with a given letter) in a fixed time (usually 60 or 90 seconds). Given the language and memory difficulties present in the population of very preterm children, we expected to find differences in their semantic memory organization compared to full-term children. More specifically and based on the existing literature, we hypothesized (1) higher ASPL, (2) lower CC and (3) higher Q. Differences in their semantic memory structure may point to differences in learning and encoding. This could help to adapt current teaching and specialized educational support, as some pedagogical practices are known to foster a more flexible semantic network [24].

Ethics

The present study is part of a larger project titled Long-term impact of early nutritional and pain management in very preterm infants on brain health and function and approved by the University ethical committee (CER-VD, protocol no. 2019-01056). All methods were carried out in accordance with the STROBE 2007 guidelines and with the principle of the Declaration of Helsinki. Each parent signed informed consent for their child, and oral consent from each child was obtained.

Design of the study

The present study was nested in an observational cohort study of very preterm children who were recruited from birth at the Clinic of Neonatology of the University Hospital of Lausanne between 2011 and 2013. They were iteratively followed up at the Development Unit of the same hospital. Further details on this cohort study have been previously described [25].

For comparison, we used a reference group of full-term children of the same age. They were recruited by one of the authors in the French-speaking side of Switzerland between 2018 and 2021 as part of a parallel study about schooling and neurocognitive development [24].

Participants

Very preterm children

The initial cohort comprised 51 very preterm neonates. Throughout the study, 2 (3.9%) patients died during the follow-up, 2 (3.9%) were excluded due to complex social situations, 2 (3.9%) were lost at previous appointments, 2 (3.9%) were unreachable and 2 (3.9%) refused to participate. Finally, 3 (5.9%) children were not able to perform the verbal fluency task (for details see S1 Fig).

Therefore, in total, 38 (75% of the initial cohort) very preterm children born between 25 and 31 gestational weeks and aged 8-10 years took part in the present study. They were tested separately in quiet rooms during their follow-up appointment at the Developmental Unit between October 2019 and October 2021.

Full-term children: reference group

Thirty-eight full-term children aged 7-10 years were matched based on age and gender. They performed different tests during their participation in the parallel study [24] and were tested at the same hospital or at their school.

Materials

Verbal fluency

The very preterm children performed the Word Generation task, a subtest of the NEPSY-II [26], a neuropsychological battery where children must generate words. It is divided into two sub-categories, the Semantic Word Generation score and the Initial Letter Word Generation score. In the Semantic Word Generation, children had sixty seconds to name all the words belonging to the category (i.e., animals and foods/drinks) they could think of. Then, in the Initial Letter Word Generation, they were given a letter (i.e., F and S) and had to produce all the words starting with this letter they could in sixty seconds. Words were immediately transcribed. The total number of correctly generated words was used. This subtest also comprises a comparison score calculated based on the scores of the two subcategories. It is aimed to compare fundamental processes (i.e., semantic) and more complex ones (i.e., initial letter).

The full-term children did not complete the whole NEPSY but only performed a verbal fluency task about animals.

Based on previous research [24,27], we selected the animal category to compute the semantic networks and compare the two groups.

Socioeconomic status (SES)

Regarding the very preterm children, SES was assessed through the Largo [28]. It is composed of two scores, the mother’s education (or current work if applicable), and the father’s current job, both scoring from 1 to 6 (1 representing higher SES). Full-term children’s SES was determined by both parent’s educational level and current job (both scoring from 1 to 4, 1 representing lower SES). These scores were summed and averaged between both parents. To compute the comparison between both groups, the very preterm score was reversed, and all scores were set as a percentage.

Non-verbal intelligence

The Matrix Reasoning subtest from the Wechsler Intelligence Scale for Children 5^th Edition (WISC V) [29] was used to measure the intelligence of very preterm children. They were asked to choose the missing piece that completed the sequence of coloured visual pattern matrices. Full-term children filled the black and white Raven's Progressive Matrices [30]. They were also asked to select the missing piece of black and white visual pattern matrices. The two scores were converted in success rate.

Data analyses

Word generation

All three scores (i.e., Semantic Word Generation, Initial Letter Word Generation, and Comparison scores) were expressed as standard scores ranging from 1 to 19. Comparison score is calculated based on Semantic Word Generation and Initial Letter Word Generation standard scores. Lower Comparison scores are representative of a profile with good capacity for vocabulary production and less effective strategies for retrieving non-categorical information, while the highest scores are the opposite.

Group comparison

The socioeconomic background and intelligence having a significant impact on verbal abilities, both variables were tested between groups using t-tests to assess group comparability. We then used them as covariates in two adjusted semantic network analyses (i.e., a first adjusted model with the SES as covariate and a second adjusted model with SES and non-verbal intelligence as covariates).

To ensure the validity of the use of the Matrix Reasoning subtest success rate for very preterm children, Pearson’s correlation was made between success rates and the Full Scale IQ T-scores from the WISC-V. We found strong correlations between both scores (r (36) = 0.729, p<.001) allowing us to use the success rate as a measure of non-verbal intelligence. We kept the Matrix Reasoning subtest to be as close as possible to the Matrix task done by the full-term children.

Semantic network analysis

The semantic network analysis was computed on the animal category. The network is decomposed in nodes and edges. Each node represents a category exemplar (i.e., in the present study an animal). Edges are the associations between two category exemplars. They represent the capacity of the groups to generate the second exemplar when they have produced the first one (e.g., generate both lion and tiger).

The fluency data were preprocessed and the semantic network components were measured using the public-available SemNA pipeline in R [31] using the R packages SemNetCleaner, SemNetDictionaries, and SemNet [31], NetworkToolbox [32], readxl [33], RDS [34],

Network estimation. The data were first pre-processed and cleaned with the SemNetDictionaries [35] and SemNetCleaner [36] tools. Repeated answers, non-categorical (i.e., others than animals), and fictional animals were excluded. Misspelled words and root variations (e.g., plural or adults and their offspring) were corrected and changed to their root (respectively, e.g., cats è cat). A binary response matrix was constructed with these cleaned responses. Its columns represent all the different exemplars given by the sample and its rows represent participants. This matrix is filled out by 1 (if an exemplar was generated by that participant) and 0 (if that exemplar was not).

To compare the two groups, an independent t-test was run on the average total number of these cleaned responses given by each child and chi-squared test on the total number of unique cleaned responses of both groups (i.e., all different words generated by at least one participant in each group.

Then, we used a bootstrapping approach to simulate and compare semantic networks of both groups. We applied the case-wise bootstrap method, meaning that for each group, we resampled with replacement from the participants in their respective groups. Each bootstrap sample has as many participants as the original sample but with some participants being included more than once and others not being included at all. For the two adjusted models, we resampled the same participants' covariate measures (i.e., SES and non-verbal intelligence) with each bootstrap sample. The bootstrap handled finalizing (i.e., we used only responses given by at least 2 participants) and equating (i.e., networks of both groups in each sample are compared using the same nodes [37]) with each replicate sample. We used the similarity to avoid negative associations between nodes in the network (i.e., all values are positive ranging from 0 (two responses do not co-occur) to 1 (two responses always co-occur)). In addition, we applied the Triangulated Maximally Filtered Graph (TMFG; [38]) as a filtering method on the word association matrix. It minimizes the noise and potential spurious associations by constructing a sub-network to capture only the most relevant information (i.e., removal of spurious associations and keeping the highest correlations) within the original network. The TMFG method keeps the number of edges the same between groups, therefore newly created networks will be comparable as having the same number of nodes and edges. This eliminates the risk of confounding different network structures being due to a different number of edges [39,40].

Networks were thus estimated and their measures (i.e., ASPL, CC, Q) were computed. This was repeated iteratively for 1,000 times.

Network analyses. We computed the 3 network coefficients for both groups (see Fig 1 for visualization). The Average Shortest Path Length (ASPL) refers to the average shortest number of steps (i.e., edges) we need to get between any pair of nodes in the network. It measures the global organization. While higher ASPL means that the semantic network is more spread out (i.e., less interconnected) and people are less likely to make associations due to greater distances between associations [41]. Lower ASPL represents a more interconnected network with shorter distances between associations. Thus, lower ASPL increases the possibility of reaching more distant associations faster, like in creativity [22,42]. The Clustering Coefficient (CC) refers to the extent that two neighbours of a node will be neighbours themselves. It illustrates the manner in which the semantic network is organized at a local level [31]. The higher this coefficient is, the more the nodes that are near neighbours to each other will tend to also be connected and the network will be more interconnected [22]. The third coefficient, the Modularity Coefficient (Q) represents the extent to which the network can be compartmentalized into small sub-networks. A sub-network can represent an animal category (e.g., sea animals, savannah animals, farm animals, pets). The higher the coefficient, the more connections there are within a sub-network and the fewer there are with nodes in other subnetworks making sub-networks more distinct from each other. It illustrates vocabulary enrichment.

Fig 1. Visual representation of the three semantic coefficients for the animal category, -i.e., the Shortest Path Length (SPL), the Clustering Coefficient (CC) and the Modularity (Q). The ASPL cannot be visually represented but is the average of all SPL of the network. Note. Adapted from “Structural differences of the semantic network in adolescents with intellectual disability”, by K. Nilsson, L. Palmqvist, M. Ivarsson, A. Danielsson, M. Annell, D. Schöld, & M. Socher, 2021, Big Data and Cognitive Computing, 5(2), p.2 (https://doi.org/10.3390/bdcc5020025) with permission. Copyright 2021 by the authors.

Then we first ran an unadjusted model and applied a t-test on the three networks coefficients (i.e., ASPL, CC, Q) to examine the difference across both groups. Thanks to a specially adapted version of the code (i.e., adapted version of the pipeline from [31]), we then ran two adjusted models. We applied a first ANCOVA with the SES as a covariate and a second with the SES and the non-verbal intelligence as covariates.

Bootstrapping small samples can result in overrepresentation of outliers. Given the relatively small size of our groups, these latter could generate significant variation in our analyses. Therefore, we ran the three bootstrapping analyses twenty times and reported the proportion of times ASPL, CC, and Q were significant to demonstrate consistency in the results and that they were not the result of a single chance event.

As the visualization of the semantic network cannot be pooled over the twenty tests and cannot be computed with the addition of the covariates, a 2D visualisation was produced for one of the twenty tests of the unadjusted model to illustrate the results. In addition, bar plots of the coefficients average of the 20 tests for the unadjusted and the two adjusted models are provided in the Supplementary Material to compare the two groups.

Demographics of very preterm children

The very preterm children from our cohort were born between 25 and 31 gestational weeks (M = 27.1, SD = 1.48). Their birth weight was between 517 and 1590 grams (M = 908, SD = 248). Comparing to the general population (M = 100, SD = 15), their general IQ assessed with the WISC was in the norms (M = 103.61, SD = 11.31). Twelve (31.58%) of them were undergoing or had undergone speech therapy.

Word generation of the very preterm children

Very preterm children performed in the norms for the semantic score with a potential deficit for the initial letter score with a score in the low norm (see Table 1 for details). The lower comparison score (M=6.95, SD=2.24) indicated a good capacity for vocabulary production but less effective strategies for retrieving non-categorical information.

Table 1. Results of the Word Generation subtest in very preterm children.

	Very preterm Children		Norms
	M(SD)	min, max	Norms
Semantic	10.4 (2.54)	5-15	10 (3)
Initial Letter	7.22 (1.95)	4-11	10 (3)
Comparison	6.95 (2.24)	3-11	/

The scores were expressed as standard scores. There are no norms for the comparison score, but lower scores represent good capacities for vocabulary production with less effective strategies for retrieving non-categorical information, while the highest scores reflect the opposite pattern.

Group comparison

Very preterm and full-term children demographics were assessed to establish their comparability (see Table 2 for details).

Table 2. Demographic and non-verbal Intelligence comparison between very preterm and full-term groups.

	Group
	Very preterm children	Full-term children	t-test or ꭓ²	p value	Cohen’s d or φ
N (% girls)	38 (52.63%)	38 (52.63%)	0	1	0
Age in years	8.78 (0.458)	8.82 (0.715)	0.32	0.75	/
Socioeconomic status	58.6 (25.1)	59 (18.1)	0.07	0.94	0.02
Non-verbal Intelligence	51.8 (11.1)	88.7 (9.82)	73	<.001	/

Semantic network analysis

Verbal fluency

Overall, very preterm children showed lower quantitative semantic abilities compared to full-term children (see Table 3 for details).

Table 3. Word generation scores separated by group.

	Group
	Very Preterm Children	Full-term Children	Welch’s t-test or ꭓ²	p value	Cohen’s d or φ
Mean total number of responses (Mean, SD)	13.05 (3.48)	17.92 (5.16)	4.82	<.001	/
Unique responses (N)	136/242	202/242	37.7	<.001	0.39
Responses produced only by one child (N)	57	96	/	/	/
Responses not provided by the other group (N)	40	106	/	/	/

Network coefficients – unadjusted model

Twenty case-wise bootstrap network analyses demonstrated that despite small samples, very preterm children had a longer Average Shortest Path (M_ASPL,VPT = 3.265 vs M_ASPL,FT = 3.222) and a lower Clustering Coefficient (M_CC,VPT = 0.697 vs M_CC,FT = 0.700) than full-term children in every tests. In contrast, regarding the Modularity Coefficient, there was no significant difference (2 times out of 20) between very preterm children (M_Q,VPT = 0.596) and full-term children (M_Q,FT = 0.597) (for the t-test results and the bar plots of the twenty case-wise bootstrap analyses, see the S2 Doc and see Fig 2 for a visual representation).

Fig 2. Visualisation of the semantic networks of the very preterm and full-term children from one of the twenty bootstrap analyses.

Network coefficients – adjusted models

When the SES was added as a covariate, the twenty case-wise bootstrap network analyses also demonstrated that very preterm children presented a longer Average Shortest Path Length (and M_ASPL,VPT= 3.265 vs M_ASPL,FT= 3.226) in 19 tests out of 20 and a lower Clustering Coefficient than full-term children (M_CC,VPT = 0.697 vs M_CC,FT= 0.700) in every 20 tests. Regarding the Modularity Coefficient, there was no significant difference (only 4 times out of 20) between very preterm children (M_Q = 0.595) and full-term children (M_Q = 0.597) (for the ANCOVA results and the bar plots of the twenty case-wise bootstrap analyses, see the S3 Doc).

When both SES and non-verbal intelligence were added as covariates, the analyses showed significant differences only regarding the Clustering Coefficient, very preterm children (M_CC = 0.693) having a lower Clustering Coefficient than full-term children (M_CC = 0.705), 16 times out of 20. Indeed, despite very preterm children (M_ASPL = 3.35) had longer Average Shortest Path Length than full-term children (M_ASPL = 3.14), it was significant in only 7 tests out of 20. As in previous analyses, the Modularity Coefficient, also demonstrated no significant difference (only 4 times out of 20) between very preterm children (M_Q = 0.602) and full-term children (M_Q = 0.591) (for the ANCOVA results and the bar plots of the twenty case-wise bootstrap analyses, see the S4 Doc).

Schoolchildren born very preterm are known to present language weaknesses, measured through quantitative tools. However, qualitative aspects of language are less investigated. In the present study, we examined the organization of concepts in verbal memory of very preterm schoolchildren. Using a verbal fluency task, we depicted their semantic network (i.e., the encoding and storage of concepts) and compared it with the semantic network of schoolchildren born at term, based on regular network metrics. Overall, very preterm children presented longer ASPL (i.e., distance between concepts) and lower CC (i.e., local interconnectivity) than full-term children. However, similar Q (i.e., compartmentalization into small sub-networks) was observed. Here, we discuss this finding through the lens of known processing speed and language use in very preterm children and how pedagogical practices could be seen as a preventive action specifically addressing this issue.

Very preterm schoolchildren are known to have language impairments, impacting semantics, expressive vocabulary, verbal language memory, and grammar [10,13]. While 12 very preterm children of our cohort were undergoing speech therapy, they overall performed in the norms at the Word Generation subtest. However, they presented subtle difficulties in retrieving information when not organized into categories, with a consequently low comparison score. It may reflect either weaknesses in search strategies due to poorer working memory and inhibition capacity (i.e., executive functions) or a deficit in orthographic fluency or sound awareness. In other words, they have encoded and stored semantic representations in long-term memory and have access to them, but they have not stored orthographic/sound representations or have difficulty retrieving them. However, in terms of categories, very preterm children generated fewer animal names, less unique ones, and poorer diversity compared to the full-term children in the present study. Consequently, if we consider only quantitative measures, we might think that very premature children have vocabulary difficulties and therefore suggest educational support based on learning additional vocabulary. Yet, the underlying difficulties might lie elsewhere. While their vocabulary enrichment (Q) was similar to those of full-term schoolchildren in the present study, the semantic network of very preterm children was less interconnected at a global (ASPL) and a local (CC) level than observed in full-term children. Given that more distant and less embedded concepts will be harder to be retrieved and used concomitantly with other concepts [17], this qualitative perspective provides new understandings on their linguistic and cognitive difficulties, namely the way they organize and connect information. Establishing associations with distant concepts is more difficult, and it is harder to retrieve information when combined with a less organized network. We can observe this difficulty in young childhood, at 18 months, as very preterm infants take more time to respond to spoken demands [43]. These authors linked this slowness to the amount of vocabulary the children had independently of their prematurity status. However, it can be hypothesized that this could also be related to the semantic network organization and the difficulty for children to access information. Across development, this may have a cascading effect on learning processes [44]. In fact, very preterm children have a lower processing speed in various tasks (e.g., [45]). This could reflect an inadequate semantic network organization with the difficulty to relate the information received to known concepts, as well as to retrieve and select concept to respond. Furthermore, while very preterm children show memory impairments in various domains (for a review see Anderson [46]), there is evidence that a less interconnected word will be less easily recalled during a memory cued recall task [47]. Therefore, their less organized network could be partly responsible for their neurodevelopmental deficits.

Very preterm children are known to experience multisensory difficulties [48,49], that are responsible for higher-level cognitive impairments (for a review see Wallace et al. [50]) including language and memory [51]. Therefore, their inability to organise new concepts optimally might be partly explained by multisensory difficulties during the learning process. In view of this vulnerability, it might be appropriate to integrate interventions promoting multi-sensoriality into the learning process. Pedagogies using didactic materials engaging multiple senses, in active-based form, such as the Montessori pedagogy [52], were found to positively impact semantic networks’ structure. Indeed, Montessori schoolchildren compared with traditionally-schooled children, have a higher CC and lower ASPL and Q [24]. Therefore, it might be relevant to see pedagogy as a preventive measure for very preterm children. Providing access to educational settings and learning materials that reinforce and train complementary sensory channels will foster qualitative language abilities. Focusing more on meaning and integrating new words while engaging multiple senses rather than teaching isolated concepts, might allow them to build a more interconnected and organized network.

However, it should be noted that when both the SES and the non-verbal intelligence were added as covariates, very preterm children only demonstrated a lower CC. Given the importance of intelligence and SES on language development, differences in semantic processes may be partly underpinned by differences in non-verbal intelligence and SES, with the caveat that different measures of non-verbal intelligence were used for the two groups, which may have impacted the results.

Some limitations need to be acknowledged. First, the experimental settings for the very preterm and full-term children were not similar. For the former, data were collected in the framework of a routine clinical assessment in the context of a longitudinal study, while for the latter, data were collected in the framework of a voluntary-based experiment. Consequently, contexts may not be similarly stressful. Second, the very preterm children constituted a heterogeneous group (i.e., gestational age, birth weight, or speech therapy), which could have introduced some confounding effects. Future studies might want to test more homogeneous groups to better understand the specific implications of the different characteristics of very preterm children. Third, bootstrapping in small sample sizes being severely affected by outliers, given our modest sample-size, the analysis was not completely robust. Nevertheless, our results showed some consistency suggesting that although both groups were small, there did not appear to be issues related to outliers. Yet, it would be relevant to reproduce this study with a bigger cohort to confirm these results. Fourth, we only evaluated animal-related semantic network. It could be interesting to repeat this task with other semantic categories, such as food or hobbies. In fact, due to its advantages (e.g., generational, cultural, linguistically independent), the semantic category of animals is the most widely used [53]. But depending on the child's interest in the category, the organisation of the network may change. Fifth, we acknowledge the non-verbal intelligence measure was not the same for both groups, nevertheless, we took the most similar measure and attested to its validity for the very preterm children. But future studies might want to use the same measure to ascertain the impact of intelligence on semantic processes. Finally, we only have a cross-sectional perspective. As longitudinal studies show that the clustering coefficient diminishes with age in healthy population [54], it would be interesting to see if it is also the case with very preterm cohorts. In addition, given that very preterm children catch up with some of their weaknesses in adolescence (i.e., phonological processing, sight word reading and linguistic working memory, as mentioned above) [9,16], we could investigate whether their semantic networks also normalize or follow the same pattern as those of full-term children while keeping these differences.

In conclusion, this study is one of the first to highlight qualitative aspects of language organization of very preterm children compared to full-term children, revealing subtle impairments. In fact, their semantic network appeared to be less connected, probably related to their processing speed and language difficulties. Given the importance of semantic network structure in higher cognitive functions, these findings encourage larger, longitudinal studies in this field. This will lead to improved support and teaching practices for very preterm children.

Acknowledgments

We would like to thank all the motivated children and parents who took part in the main longitudinal study and who have agreed to take part in this final stage, thus contributing to the realization of various studies. We also thank Dr. Jérôme Rime who has helped us with the interpretation of the NEPSY results.

Author contributions

MD: conceptualization, data curation, formal analysis, investigation, methodology, visualization, writing-original draft preparation

APC: software, methodology, writing – review & editing

LB: investigation, data curation, writing – review & editing

CHV: investigation, data curation, writing – review & editing

CJFF: funding acquisition, project administration, writing – review & editing

SD: conceptualization, data curation, investigation, methodology, writing – review & editing

JS: conceptualization, funding acquisition, methodology, project administration, supervision, writing – review & editing

Competing interests

The authors declare that they have no known competing financial interests or personal relationships that are relevant to the content of this article.

Funding

This work was supported by the Fondation W. et E. Grand d’Hauteville in Lausanne (Switzerland).

Data availability

The anonymised experimental data collected, the specific xlsx files for analysis, and the analytic code written in R are publicly available on the Open Science Framework repository (available once published).

1. Doyle LW, Spittle A, Anderson PJ, Cheong JLY. School-aged neurodevelopmental outcomes for children born extremely preterm. Arch Dis Child. 2021;106(9):834–8.

2. Marlow N, Ni Y, Lancaster R, Suonpera E, Bernardi M, Fahy A, et al. No change in neurodevelopment at 11 years after extremely preterm birth. Arch Dis Child-Fetal Neonatal Ed. 2021;106(4):418–24.

3. van Baar AL, van Wassenaer AG, Briët JM, Dekker FW, Kok JH. Very Preterm Birth is Associated with Disabilities in Multiple Developmental Domains. J Pediatr Psychol. 2005 Apr 1;30(3):247–55.

4. Woodward LJ, Moor S, Hood KM, Champion PR, Foster-Cohen S, Inder TE, et al. Very preterm children show impairments across multiple neurodevelopmental domains by age 4 years. Arch Dis Child-Fetal Neonatal Ed. 2009;94(5):339–44.

5. Luu TM, Ment LR, Schneider KC, Katz KH, Allan WC, Vohr BR. Lasting Effects of Preterm Birth and Neonatal Brain Hemorrhage at 12 Years of Age. Pediatrics. 2009 Mar 1;123(3):1037–44.

6. Vandormael C, Schoenhals L, Hüppi PS, Filippa M, Borradori Tolsa C. Language in Preterm Born Children: Atypical Development and Effects of Early Interventions on Neuroplasticity. Neural Plast. 2019;2019:6873270.

7. van Noort-van der Spek IL, Franken MCJP, Weisglas-Kuperus N. Language Functions in Preterm-Born Children: A Systematic Review and Meta-analysis. Pediatrics. 2012 Apr 1;129(4):745–54.

8. Pérez-Pereira M. Prevalence of Language Delay among Healthy Preterm Children, Language Outcomes and Predictive Factors. Child Basel Switz. 2021 Apr 6;8(4):282.

9. Luu TM, Vohr BR, Allan W, Schneider KC, Ment LR. Evidence for Catch-up in Cognition and Receptive Vocabulary Among Adolescents Born Very Preterm. Pediatrics. 2011 Aug 1;128(2):313–22.

10. Vohr B. Speech and language outcomes of very preterm infants. Semin Fetal Neonatal Med. 2014 Apr 1;19(2):78–83.

11. Kunnari S, Yliherva A, Paavola L, Peltoniemi OM. Expressive Language Skills in Finnish Two-Year-Old Extremely- and Very-Low-Birth-Weight Preterm Children. Folia Phoniatr Logop. 2012;64(1):5–11.

12. Sanchez K, Spittle AJ, Boyce JO, Leembruggen L, Mantelos A, Mills S, et al. Conversational Language in 3-Year-Old Children Born Very Preterm and at Term. J Speech Lang Hear Res JSLHR. 2020 Jan;63(1):206–15.

13. Guarini A, Sansavini A, Fabbri C, Savini S, Alessandroni R, Faldella G, et al. Long-term effects of preterm birth on language and literacy at eight years. J Child Lang. 2010;37(4):865–85.

14. Barre N, Morgan A, Doyle LW, Anderson PJ. Language abilities in children who were very preterm and/or very low birth weight: a meta-analysis. J Pediatr. 2011 May;158(5):766-774.e1.

15. Omizzolo C, Scratch SE, Stargatt R, Kidokoro H, Thompson DK, Lee KJ, et al. Neonatal brain abnormalities and memory and learning outcomes at 7 years in children born very preterm. Memory. 2014 Aug 18;22(6):605–15.

16. Lee ES a, Yeatman JD b, Luna B c, Feldman HM a. Specific language and reading skills in school-aged children and adolescents are associated with prematurity after controlling for IQ. Neuropsychologia. 2011 Apr;49(5):906–13.

17. Gobbo C, Chi M. How knowledge is structured and used by expert and novice children. Cogn Dev. 1986;1(3):221–37.

18. Biemiller A, Slonim N. Estimating root word vocabulary growth in normative and advantaged populations: Evidence for a common sequence of vocabulary acquisition. J Educ Psychol. 2001;93(3):498.

19. Tulving E. Episodic and semantic memory. In: Organization of memory. Oxford, England: Academic Press; 1972. p. xiii, 423–xiii, 423.

20. Markovits H, Fleury ML, Quinn S, Venet M. The development of conditional reasoning and the structure of semantic memory. Child Dev. 1998;69(3):742–55.

21. Vales C, States SL, Fisher AV. Experience-Driven Semantic Differentiation: Effects of a Naturalistic Experience on Within- and Across-Domain Differentiation in Children. Child Dev. 2020;91(3):733–42.

22. Siew CSQ, Wulff DU, Beckage NM, Kenett YN. Cognitive Network Science: A Review of Research on Cognition through the Lens of Network Representations, Processes, and Dynamics. Complexity. 2019 Jun 17;2019:e2108423.

23. Goñi J, Martincorena I, Corominas-Murtra B, Arrondo G, Ardanza-Trevijano S, Villoslada P. Switcher-random-walks: A cognitive-inspired mechanism for network exploration. Int J Bifurc Chaos. 2010;20(03):913–22.

24. Denervaud S, Christensen AP, Kenett YN, Beaty RE. Education shapes the structure of semantic memory and impacts creative thinking. Npj Sci Learn. 2021 Dec 9;6(1):1–7.

25. Schneider J, Kober T, Bickle Graz M, Meuli R, Hüppi PS, Hagmann P, et al. Evolution of T1 Relaxation, ADC, and Fractional Anisotropy during Early Brain Maturation: A Serial Imaging Study on Preterm Infants. Am J Neuroradiol. 2016 Jan;37(1):155–62.

26. Korkman M, Kirk U, Kemp S. NEPSY, bilan neuropsychologique de l’enfant [adaptation française 2003]. Montreuil Ed ECPA. 1997;

27. Barker JE, Semenov AD, Michaelson L, Provan LS, Snyder HR, Munakata Y. Less-structured time in children’s daily lives predicts self-directed executive functioning. Front Psychol [Internet]. 2014 [cited 2022 May 24];5. Available from: https://www.frontiersin.org/article/10.3389/fpsyg.2014.00593

28. Largo RH, Pfister D, Molinari L, Kundu S, Lipp A, Due G. Significance of prenatal, perinatal and postnatal factors in the development of AGA preterm infants at five to seven years. Dev Med Child Neurol. 1989;31(4):440–56.

29. Wechsler D. WISC-V: Manuel technique et d’interprétation. PsychCorp Ed Don Mills Pearson. 2016;

30. Raven J, Raven JC, Court JH. Manual for Raven’s progressive matrices and vocabulary scales. Section 1: General overview. Harcourt Assessment. 2003;

31. Christensen AP, Kenett YN. Semantic network analysis (SemNA): A tutorial on preprocessing, estimating, and analyzing semantic networks. Psychol Methods. 2021;

32. Christensen AP. NetworkToolbox: Methods and Measures for Brain, Cognitive, and Psychometric Network Analysis in R. R J. 2018;10(2):422–39.

33. Wickham H, Bryan J. readxl: Read excel files. R Package Version. 2019;1(1):785.

34. Handcock MS, Fellows IE, Gile KJ. RDS: Respondent-driven sampling. Version 07-7. 2016;

35. Christensen AP. SemNetDictionaries: Dictionaries for the’SemNetCleaner’package. 2019.

36. Christensen AP. SemNetCleaner: An automated cleaning tool for semantic and linguistic data. Package Version. 2019;1(0).

37. Kenett YN, Wechsler-Kashi D, Kenett DY, Schwartz RG, Ben-Jacob E, Faust M. Semantic organization in children with cochlear implants: Computational analysis of verbal fluency. Front Psychol. 2013;4:543.

38. Massara GP, Di Matteo T, Aste T. Network Filtering for Big Data: Triangulated Maximally Filtered Graph [Internet]. arXiv; 2015 [cited 2022 Sep 5]. Available from: http://arxiv.org/abs/1505.02445

39. Borodkin K, Kenett YN, Faust M, Mashal N. When pumpkin is closer to onion than to squash: The structure of the second language lexicon. Cognition. 2016;156:60–70.

40. Christensen AP, Kenett YN, Aste T, Silvia PJ, Kwapil TR. Network structure of the Wisconsin Schizotypy Scales–Short Forms: Examining psychometric network filtering approaches. Behav Res Methods. 2018 Dec 1;50(6):2531–50.

41. Christensen AP, Kenett YN, Cotter KN, Beaty RE, Silvia PJ. Remotely Close Associations: Openness to Experience and Semantic Memory Structure. Eur J Personal. 2018 Jul 1;32(4):480–92.

42. Benedek M, Kenett YN, Umdasch K, Anaki D, Faust M, Neubauer AC. How semantic memory structure and intelligence contribute to creative thought: a network science approach. Think Reason. 2017 Apr 3;23(2):158–83.

43. Marchman VA, Ashland MD, Loi EC, Adams KA, Fernald A, Feldman HM. Predictors of early vocabulary growth in children born preterm and full term: A study of processing speed and medical complications. Child Neuropsychol. 2019 Oct 3;25(7):943–63.

44. Rose SA, Feldman JF, Jankowski JJ. Modeling a cascade of effects: The role of speed and executive functioning in preterm/full-term differences in academic achievement. Dev Sci. 2011;14(5):1161–75.

45. Brydges CR, Landes JK, Reid CL, Campbell C, French N, Anderson M. Cognitive outcomes in children and adolescents born very preterm: a meta-analysis. Dev Med Child Neurol. 2018;60(5):452–68.

46. Anderson PJ. Neuropsychological outcomes of children born very preterm. Semin Fetal Neonatal Med. 2014 Apr 1;19(2):90–6.

47. Nelson DL, Bennett DJ, Gee NR, Schreiber TA, McKinney VM. Implicit memory: Effects of network size and interconnectivity on cued recall. J Exp Psychol Learn Mem Cogn. 19931201;19(4):747.

48. Bröring T, Oostrom KJ, Lafeber HN, Jansma EP, Oosterlaan J. Sensory modulation in preterm children: Theoretical perspective and systematic review. Key A, editor. PLOS ONE. 2017 Feb 9;12(2):e0170828.

49. Bröring T, Königs M, Oostrom KJ, Lafeber HN, Brugman A, Oosterlaan J. Sensory processing difficulties in school-age children born very preterm: An exploratory study. Early Hum Dev. 2018 Feb 1;117:22–31.

50. Wallace MT, Woynaroski TG, Stevenson RA. Multisensory integration as a window into orderly and disrupted cognition and communication. Annu Rev Psychol. 2020;71:193–219.

51. Quak M, London RE, Talsma D. A multisensory perspective of working memory. Front Hum Neurosci [Internet]. 2015 [cited 2023 May 30];9. Available from: https://www.frontiersin.org/articles/10.3389/fnhum.2015.00197

52. Marshall C. Montessori education: a review of the evidence base. Npj Sci Learn. 2017 Oct 27;2(1):1–9.

53. Ardila A, Ostrosky-Solís F, Bernal B. Cognitive testing toward the future: The example of Semantic Verbal Fluency (ANIMALS). Int J Psychol. 2006;41(5):324–32.

54. Dubossarsky H, De Deyne S, Hills T. Association networks across the lifespan: the large-scale structure of free association networks, evidence for U-shaped developmental changes and semantic saturation. 2014. S38 p.

The authors declare no competing interests.

S1Fig.tif
Flowchart of the very preterm and full-term children
S2doc.docx
S3doc.docx
S4doc.docx

Download PDF

Version 2

posted

You are reading this latest preprint version

Characterization of language abilities and semantic networks in very preterm children at school-age

Status:

Version 2

Abstract

Figures

Introduction

Materials And Methods

Results

Discussion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 2