A total of 97 children with dyslexia participated in the study, selected from a larger sample of children recruited from the database of the Dyslexia Center at UCSF (see inclusion/exclusion criteria below), and were part of multidisciplinary research program that performs neurological, psychiatric, cognitive and language evaluations of children with Developmental Dyslexia (DD). Participants underwent comprehensive academic, cognitive and language assessments, and were given a diagnosis of DD based on International Dyslexia Association (IDA; Lyon et al., 2003; Definition Consensus Project) definition. All met inclusion criteria: (1) native speakers of English language, (2) 8–13 years of age, (3) intelligence estimates within normal limits (WASI Matrix Reasoning or Receptive One-Word Picture Vocabulary >5th percentile; Wechsler, 1999), and (4) adequate sensorimotor capacity, including normal or corrected-to-normal vision and hearing. Reading ability was measured by using standardized age-adjusted measures. A participant was excluded from the dyslexia cohort if all their reading scores fell above the 25^th percentile: isolated timed word and pseudoword reading (TOWRE-2 Sight Word Efficiency and Phonemic Decoding Efficiency; Torgesen et al., 2012), untimed word and pseudoword reading (Woodcock Johnson Letter Word Identification and Word Attack; Schrank et al., 2014), and/or paragraph reading (GORT measures; Wiederholt and Bryant, 2012). All children were screened from schools in Northern California.

On average, percentile performance across fluency tests ranged from 2 to 99.99 percentiles, with mean percentile performance of 77.3 (SD = 23.85) on semantic fluency and 34.67 (SD = 24.85) on letter fluency across groups. The median age-normed percentile and interquartile range for the fluency tests are displayed in Table 1, separately for each group. The two semantic groups were matched on letter fluency, matrix reasoning, and vocabulary.

Both high and average groups were highly accurate on CVLT-C Trial 1, with a mean percent correct accuracy of 60.8 ± 27.8 and 45.6 ± 30.2, respectively. However, the high and average groups were less accurate on the Children’s Colored Trails B test, with a mean percent correct accuracy of 28.3 ± 20.4 and 28.3 ± 18.9, respectively. We found a significant main effect of group on CVLT-C Trial 1 [F_(1,94) = 6.193, p = 0.014, Cohen’s f = 0.26] and not on Children’s Colored Trails B [F_(1,91) = 2.292, p = 0.133, Cohen’s f = 0.16], but this did not survive correction for multiple comparisons.

We had item-by-item data on the semantic fluency test for 77 of the initial 97 participants (see Supplementary Table S1 on 53 children in high and 24 children in average semantic fluency group); these participants’ data were included in clustering and switching analyses. This subgroup did not differ meaningfully from the full sample on demographic variables. See Supplementary Table S1 on statistics of demographics between the two groups selected for clustering and switching analyses.

The high semantic fluency group demonstrated different internal states of the brain (i.e., spatial features from the dynamic functional connectivity) compared to the average semantic group (p < =0.05/246, where 246 is the number of ROIs) in the left inferior and middle frontal gyri (IFG/MFG), posterior inferior temporal gyri (ITG), and dorsal medial parietal lobe as well as right IFG, ITG, inferior parietal, temporoparietal junction, and occipital regions (see Figure 3). The results also revealed that the high semantic fluency group exhibits a significantly shorter average dwell time in this network of regions, suggesting that their brain states change more frequently compared to the average semantic fluency group. The high semantic fluency group switches more frequently between brain states compared to the average semantic fluency group do on average (number of change points: 7.4706 vs. 5.862, respectively; p-value = 0.037). This is consistent with the observation that the maximal dwell time in a stationary time segment is significantly shorter in the high semantic fluency group than the average semantic fluency group with (52.754 vs. 83.611, respectively; p-value = 0.055). These results suggest that the high semantic fluency group is significantly more active in brain state switches. The significant ROIs demonstrating the significant dynamic rs-fMRI connectivity are listed in Supplementary Table S3.

We believe the strong performance in semantic fluency in children with dyslexia is related to semantic adeptness but isolating semantics from other cognitive skills is challenging, similar to other studies (Aita et al., 2018). However, once we regressed letter fluency on semantic fluency and grouped children based on a residual score, we were able to obtain well-matched groups aside from semantic fluency performance. The particular semantic ability identified would likely be more related to the production or selection of specific words based on lexical-semantic knowledge rather than comprehension of higher-level semantics, such as deep cohesion or referential cohesion (e.g., Dahl et al., 2020). Similarly, semantic fluency categories are usually based on high-frequency, concrete categories that may not be representative of all aspects of semantics, particularly abstract or lower-frequency categories. For example, it is common for children to participate in everyday conversations about animals or food/drink, but it is less common for them to participate in conversations about evolution or non-poisonous plant species (Williams et al., 2013). Therefore, this strength in semantics should be interpreted within the confines of the types of categories used, i.e., concrete, high frequency, and the means of examination, i.e., production.

While interpreting results, it is important to consider whether children with dyslexia have a strength in semantics or are using semantics to compensate for their difficulties with phonology. First, ironically, part of that debate is a matter of semantics in the sense that compensation compared to alternate route needs to be better defined in order to frame the debate. Two key areas to describe are the order of development and the typical distribution of work. If one ability is acquired first and strength in the other ability is only noted when the first ability is compromised, then compensation may be an appropriate term. On the other hand, if development is simultaneous and strength is normally distributed, then alternate is likely a better description. Similarly, if a particular ability requires multiple cognitive skills and there is a normal distribution of a balance of skills, then alternate would be a better term to characterize strengths. Finally, most of the studies suggesting children with dyslexia use semantics to compensate for difficulties with phonology are specifically based on reading (Levesque et al., 2017). Our semantic fluency results demonstrate that within dyslexia there is a distribution of semantic ability that is independent of phonology. Further, semantic fluency does not differentiate reading ability within our sample of youth with dyslexia, because our groups are matched on those abilities. Therefore, we favor the interpretation that, in our data, semantic fluency serves as a strength rather than a form of compensation.

Several studies have examined semantic fluency in English-speaking children with dyslexia. The study by Levin (1990) reported that children with dyslexia had lower verbal fluency, including semantic fluency, compared to age-matched controls. Levin assessed fluency in a way similar to our study (number of correct words per minute) and used a similar category (food). However, there are a few important caveats to Levin’s study: (1) they reported that the dyslexia group generated 14.1 names compared to 10.4 names generated by the control group in their Table, while the statistics reported in the text claimed the opposite. Additionally, the mean number of words generated for controls does not sum within a small rounding error even if the proper names are reversed between dyslexia and controls, which suggests there were likely data sorting errors. Thus, the actual results of the study are in question; (2) The groups were not clearly matched on general intelligence, and they did not report means and standard deviations on the measure used for an intelligence proxy (Peabody Picture Vocabulary Test), which associates with fluency tests (Friesen et al., 2021; Regard et al., 1982), and (3) only male participants were included, which may influence performance on certain fluency tests (Hirnstein et al., 2023).

A second study found that children with dyslexia had stronger semantic fluency performance, average design fluency, and lower phonemic fluency compared with controls (Griffiths, 1991). This pattern of findings is similar to the distributions of semantic, design, and letter fluency performances in our data, and the age and intelligence scores of their participants were similar to our cohort. However, results cannot be directly compared, as they used semantic cueing methods of extracting fluency, i.e., saying a word that corresponds to a cue within a certain time interval (cue: big; response: elephant) and replacing a nonsense word with a real word based on a series of five descriptive sentences. Additionally, their participants were English-speaking but Scottish, which may influence education approaches and/or vocabulary. Lastly, a third study (Frith et al., 1995) also used a different method from ours to interrogate semantic fluency and recorded how many seconds it took for participants to say 10 correct words. They found that the children with dyslexia were slower than controls to produce words based on phonemic but not semantic cues. Their participants were also older than our participants by roughly 2 years. However, despite these differences, a similar pattern was found in our study in that the children with dyslexia had difficulties with phonemic but not semantic fluency.

Notably, the children in our study primarily attended private schools that are designed to teach children with learning difficulties. Therefore, it is possible that the higher normed percentiles score we obtained on semantic fluency reflects the impact private schools have on students’ semantics rather than higher than average innate abilities in children with dyslexia (Jacobsen et al., 2017). Although our study was not designed to directly address this potential confound, there are a few variables in our dataset that provide some insights. First, the semantic fluency normative scores within our cohort are higher than vocabulary, nonverbal reasoning, and other fluency measures, which indicates relative strength in semantics within the cohort. Second, not all children in our cohort performed above average on the semantic fluency task (i.e., members of the average semantic fluency group), but they went to the same schools and had similar vocabulary, nonverbal reasoning, and family income compared to high semantic fluency participants. These ancillary observations may suggest the strength we identified cannot be entirely accounted for by private schooling.

Lastly, according to past research, semantic cognition can be broken down into two components: semantic control and semantic knowledge (Lambon Ralph et al., 2017; Vonk et al., 2020). On the one hand, semantic strength might be related to the ability to retrieve more words from memory within the semantic network (semantic knowledge for a given subcategory) and hold them in verbal short-term memory. On the other, semantic strength might be associated with higher cognitive flexibility (cognitive control processes) that support semantic memory retrieval. Our groups did not show differences in semantic knowledge as measured by the Receptive One-Word Picture Vocabulary Test. We did see some differences in executive functions subserving semantic retrieval, such as verbal short-term memory (CVLT-C Trial 1), but these differences did not withstand correction for multiple comparisons. The CVLT-C Trial 1 may be more likely than other verbal short-term memory measures (e.g., digit span) to associate with semantic cognition specifically because of its embedded semantic categories. It is possible that if someone implicitly detects the embedded structure, they may also perform better on the task than someone who does not. This is in line with other studies showing relationships between verbal fluency and short-term and working memory (Aita et al., 2018; Marques et al., 2022), but these studies were confounded by phonological impairment. The semantic cognition components are particularly relevant to our findings on clustering and switching during the semantic fluency task (as discussed below). Some children with strong semantic abilities may use strategic processes that enable them to generate words within a subcategory by leveraging their capacity to access more words from semantic memory. Additionally, they may efficiently shift to a new subcategory when the current one is exhausted, relying on semantic control processes.

To further examine the psycholinguistic correlates of the semantic fluency task, we employed a clustering and switching analysis. Traditionally, clustering and switching analysis have been used to detect subtle differences in cognitive processes across the lifespan, in healthy aging (Amunts et al., 2020; Gonzalez-Burgos et al., 2019) and in various clinical disorders including neurodegenerative diseases (van den Berg et al., 2022; Wright et al., 2023), bipolar disease (Weiner et al., 2019), schizophrenia (Barattieri di San Pietro et al., 2023), traumatic brain injury (Woods et al., 2016), and stroke (Carpenter et al., 2021). Fluency tests was found to be associated with other cognitive abilities such as creative processes (Ovando-Tellez et al., 2022). Few studies have used this analysis in children (Hurks et al., 2010; Kavé et al., 2008) and even fewer have examined this in children with dyslexia and/or developmental language disorder (see Arán Filippetti et al., 2023 for review). Although controversies exist regarding the methods of quantifying a semantic search strategy (Voorspoels et al., 2014), we chose a widely used and studied approach (Troyer, 2000).

The clustering and switching analysis revealed that the high semantic fluency group outperformed the average semantic fluency group on raw clustering and switching variables, clustering ratio, and max cluster size, which is not impacted by the total number of words generated. They produced more words overall, had a higher number of clusters and switches, and produced larger clusters of semantically related items at their max performance. The groups did not differ on the ratio of switching/total number of words produced, which suggests that, in general, both groups of children switched between clusters at a similar rate, roughly every five words.

More research is needed to determine the pattern of performance in clustering and switching components in child clinical populations. Based on a recent systematic review (Arán Filippetti et al., 2023), out of the 33 studies that were reviewed, 18 analyzed clustering and switching strategies in participants with atypical development and only three included children with dyslexia and/or developmental disorders (Mengisidou and Marshall, 2019; Mengisidou, 2020; Mielnik et al., 2015). Accordingly, different profiles are observed in atypical development, with patterns that distinguish between impairment in cognitive flexibility (less switching) and the organization of semantic and phonological representations (smaller cluster size). Similarly, our findings suggest that children with semantic strength might implement a strategic process that helps them to generate words within a subcategory (larger cluster size and higher number of clusters) and to shift efficiently to a new subcategory when a subcategory is exhausted (higher number of switches). These findings may suggest developmental changes in the structure of children’s semantic networks (Vales et al., 20,220) which may shape controlled semantic retrieval (Marko and Riečanský, 2021).

Dynamic rs-fMRI connectivity has been used to study task-free resting-state in several disease states (Yang et al., 2020). Similarly, we used dynamic rs-MRI to examine a semantic strength in dyslexia, and whether group differences were apparent in the brain’s spontaneous dynamics and spatiotemporal patterns. The results revealed that the high semantic fluency group exhibited a significantly shorter average dwell time, implying that these individuals may remain in a single brain state for shorter periods and may switch between brain states more efficiently. This study is model-based and exploratory, aimed at identifying potential explanations for group differences. Since microstates are not directly observable by humans, there is no definitive “ground truth” for the time spent in each state. The purpose of this analysis was to identify meaningful and interpretable features that could provide insights into underlying brain processes.

The dynamic rs-fMRI connectivity analysis specifically revealed a large-scale brain network potentially supporting semantic strength, including the left IFG/MFG, posterior ITG and medial parietal lobe, and the right IFG, inferior parietal regions, and occipitotemporal regions. This brain finding at rest is supported by the functional neuroimaging studies in healthy adults (see meta-analyses by Wagner et al., 2014) and lesion studies in stroke patients (Biesbroek et al., 2021) which have shown that word search during category-fluency tasks rely on the coordinated activity of several brain areas including the left IFG/MFG and posterior regions of the MTG and ITG as well as the parahippocampal and fusiform gyri. The finding is also in line with studies showing better fluency performance associated with larger surface area in brain regions related to semantic fluency in typically developing children (Gonzalez et al., 2021), and those showing specialization of executive systems (Satterthwaite et al., 2013) and semantics (Wang et al., 2021) during brain development, and is consistent with recent meta-analyses of semantic cognition in children (Enge et al., 2021). The data might suggest that higher connectivity of the so-called “semantic network” facilitates efficient information flow associated with easier switching and bigger cluster size during semantic fluency.

A key strength of our well-matched participant groups is that they were recruited from the same schools, received evidence-based reading interventions, and have similar reading scores. As a result, it is unlikely that the differences in brain dynamics are due to differences in exposure to reading instruction. Instead, it is more plausible that the differences in performance on the semantic fluency test and in the time-varying dynamic network reflect a characteristic, though potentially malleable, distinction between the semantic fluency groups. Some limitations of the study include unequal sample sizes for the high and average semantic fluency groups. It is also unfortunate that we do not have a control sample with similar socioeconomic metrics and no reading concerns to whom we could compare semantic fluency, psycholinguistic, and imaging results. Additionally, it will be important to examine semantic processing in students who are more representative of the population, i.e., diverse samples, particularly in terms of race, bilingualism, family income, and students enrolled in public schools. Finally, our imaging study was relatively small and preliminary, so larger studies that attempt to replicate the finding and also test for a continuous relationship with behavior will be valuable. Further out-of-sample validation is also required to confirm that the brain dynamics we have identified are generalizable across different studies. Given that our study was cross-sectional, it will also be important to conduct longitudinal studies that examine the change in brain dynamics before and after reading interventions. Future research would also benefit from investigating whether similar or different neurocognitive mechanisms are observed in both semantic and letter fluency tasks using the same dynamic task-free resting state approach and in both children with and without dyslexia, as this could provide deeper insights into their interplay within dyslexic populations.