Rest Profiles:
We identified four prominent subtypes of functional connectivity, each characterized by distinct patterns of mean z-scores and standard deviations (Fig XX). These subtypes were highly reproducible (0.98-0.996) based on their across-sample-matched maximum correlations, indicating reliability for further investigation (Fig XX). These subtypes are presented in separate sections, focusing on the details of the descriptions in relation to the prominent means that characterize each subtype.
Subtype 1: Auditory-Sensorimotor Integration (ASI) ASI is marked by strong connections between the auditory and sensorimotor-hand regions. This subtype is characterized by a noteworthy positive mean z-score in Subtype 3 (1.03 ± 1.10), suggesting enhanced integration between auditory processing and sensorimotor-hand regions. Conversely, Subtype 2 demonstrates a moderately negative mean z-score (-0.54 ± 0.76), indicating weaker connections in these areas.
Subtype 2: Auditory-Visual Integration (AVI) AVI is characterized by interactions between the auditory and visual regions. In this subtype, Subtype 3 displays a relatively strong positive mean z-score (0.71 ± 0.97), highlighting the presence of robust connections between auditory and visual processing areas. In contrast, Subtype 4 has a notable negative mean z-score (-0.86 ± 0.81), signifying weaker integration between these regions.
Subtype 3: Cingulo-Opercular Modulation (COM) COM encompasses two subcategories: (a) COM Default: This subcategory reveals a positive mean z-score in Subtype 2 (0.52 ± 0.84), suggesting increased connectivity within the default mode network and its integration with the cingulo-opercular network. In contrast, Subtype 1 demonstrates a negative mean z-score (-0.82 ± 0.76), pointing to weaker connections in these areas. (b) COM None: This subcategory shows a moderately positive mean z-score in both Subtype 2 (0.43 ± 0.82) and Subtype 4 (0.23 ± 0.92), indicating a more balanced connectivity pattern within the cingulo-opercular network and its interactions with other brain regions.
Subtype 4: Sensorimotor Coordination (SC) SC comprises three subcategories: (a) SC Sensorimotor Mouth-Hand: This subcategory is characterized by a prominent positive mean z-score in Subtype 3 (1.16 ± 1.04), suggesting enhanced coordination and connectivity between the sensorimotor mouth and hand regions. (b) SC Left-Caudate and Right-Hippocampus: These subcategories are characterized by negative mean z-scores in Subtype 3 (-1.08 ± 0.98 and -1.15 ± 0.92, respectively), indicating weaker connections between the cingulo-opercular network and the left caudate and right hippocampus regions. (c) SC Left-Pallidum and Right-Caudate: These subcategories exhibit similar patterns with negative mean z-scores in Subtype 3 (-1.11 ± 0.86 and -1.15 ± 0.90, respectively), signifying reduced connectivity between the sensorimotor-hand regions and the left pallidum and right caudate.
Our results demonstrate a high classification performance of the XGBoost classifier, with an overall accuracy of 0.882. The classifier exhibited remarkable precision, recall, and F1-scores across all subtypes. Specifically, subtype 1 achieved a precision of 0.906, recall of 0.873, and F1-score of 0.889. Subtype 2 displayed a precision of 0.876, recall of 0.851, and F1-score of 0.863. Subtype 3 exhibited a precision of 0.934, recall of 0.905, and F1-score of 0.919. Lastly, subtype 4 demonstrated a precision of 0.816, recall of 0.913, and F1-score of 0.862.
The SHAP values analysis provided insights into the most important features contributing to the classification performance. This model-agnostic approach allowed for a better understanding of the underlying biological mechanisms driving the classification of RSFC subtypes. The high classification performance coupled with the interpretability of SHAP values highlights the potential of our approach in elucidating the complex organization of functional connectivity patterns in the human brain.
Table XX: Classification output for subtypes
Subtype precision recall f1-score support 1 0.906 0.873 0.889
1507.000 2 0.876 0.851 0.863 1490.000 3 0.934 0.905 0.919 1168.000 4
0.816 0.913 0.862 1134.000 accuracy 0.882 0.882 0.882
macro avg 0.883 0.885 0.883 5299.000 weighted avg 0.884 0.882 0.883
5299.000
Motion Significant differences were discovered in both mean motion (p < 0.001) and maximum motion (p < 0.001) across the subtypes. Subtype 2 had the highest mean motion score of 0.44 (SD = 0.42) and the highest maximum motion score of 8.06 (SD = 7.22), while Subtype 1 had the lowest mean motion score of 0.22 (SD = 0.27) and the lowest maximum motion score of 4.26 (SD = 5.22). The pairwise comparisons of the subtypes indicated significant differences, except for Subtype 3 and Subtype 4. These findings suggest that there are distinct motion profiles among rs-fMRI subtypes and that the identified subtypes are not independent.
Cognitive, Behavioral
CBCL The resting-state fMRI subtypes were characterized based on their Child Behavior Checklist (CBCL) scores. The analyses involved a full sample and two smaller samples (Sample 1 and Sample 2) for evaluating the robustness of the identified differences. The results presented here focus on the higher-level characterization of the subtypes based on their CBCL profiles, highlighting the robust F statistics, False Discovery Rate (FDR) corrected p-values, and post-hoc comparisons. In the full sample analysis, Subtype 1 showed consistently lower scores across the majority of CBCL measures when compared to Subtypes 2 and 3. This finding suggests that Subtype 1 may be associated with fewer behavioral problems, including aggressive behavior, attention problems, externalizing behaviors, rule-breaking behaviors, and social problems. Subtype 3, on the other hand, consistently scored higher on these measures compared to Subtype 1 and, in some cases, Subtype 2. This pattern implies that Subtype 3 may be associated with more severe behavioral problems. Subtype 2 had a more varied profile, with some measures indicating higher scores than Subtype 1 but lower scores than Subtype 3. This could suggest an intermediate position for Subtype 2 in terms of the severity of behavioral problems. Subtype 4 displayed mixed results, with no consistent pattern across measures. However, some measures showed Subtype 3 having higher scores than Subtype 4. Robust findings, those identified as significant in both Sample 1 and Sample 2, were observed for attention problems (F = 10.32, FDR-corrected p < 0.001), externalizing behaviors (F = 8.21, FDR-corrected p < 0.01), rule-breaking behaviors (F = 6.98, FDR-corrected p < 0.01), and social problems (F = 5.62, FDR-corrected p < 0.01). Post-hoc comparisons revealed that Subtype 1 had significantly lower scores than Subtypes 2 and 3 on these measures, while Subtype 3 had higher scores than Subtype 2. These findings reinforce the characterization of Subtypes 1, 2, and 3, with Subtype 1 having fewer problems, Subtype 2 being intermediate, and Subtype 3 having more severe problems. In summary, the resting-state fMRI subtypes exhibit distinct profiles based on their CBCL scores. Subtype 1 is associated with fewer behavioral problems, while Subtype 3 shows more severe problems. Subtype 2 appears to be intermediate in severity. The robust findings, with significant F statistics and FDR-corrected p-values, highlight the consistency of these relationships across different samples, providing valuable insights into the potential behavioral manifestations of these subtypes. Further research is needed to understand the neural correlates of these behavioral patterns and their implications for interventions and treatment approaches.
COGNTIVE The study revealed significant differences in cognitive performance across the four subtypes in various cognitive domains. Post-hoc comparisons indicated that Subtypes 1 and 4 consistently demonstrated higher cognitive performance compared to Subtypes 2 and 3. These cognitive domains included cognitive flexibility, attention, inhibitory control, verbal learning, memory, nonverbal abstract reasoning, receptive language, and fluid intelligence. In the Flanker Inhibitory Control task, a significant difference was observed among subtypes (F(3, 102) = 45.171, p < 0.001). Subtypes 1 and 4 outperformed Subtypes 2 and 3. Similarly, the List Sorting task, measuring working memory, showed significant differences across subtypes (F(3, 104) = 40.132, p < 0.001), with Subtypes 1 and 4 outperforming Subtypes 2 and 3. The Picture Vocabulary task, assessing receptive language, also revealed significant differences among subtypes (F(3, 106) = 37.310, p < 0.001). Subtypes 1 and 4 demonstrated superior performance relative to Subtypes 2 and 3. In the Picture Sequence task, which measures episodic memory, significant differences were observed across subtypes (F(3, 103) = 24.262, p < 0.001), with Subtypes 1 and 4 outperforming Subtypes 2 and 3. For the RAVLT task, assessing verbal learning and memory, significant differences were found among subtypes (F(3, 104) = 38.305, p < 0.001), with Subtypes 1 and 4 outperforming Subtypes 2 and 3. The Pearson Matrix Reasoning task, a measure of fluid intelligence, showed significant differences across subtypes (F(3, 106) = 47.154, p < 0.001), with Subtypes 1 and 4 outperforming Subtypes 2 and 3. Lastly, in the Dimensional Change Card Sort task, measuring cognitive flexibility and set-shifting, significant differences were observed among subtypes (F(3, 104) = 50.312, p < 0.001), with Subtypes 1 and 4 outperforming Subtypes 2 and 3. In conclusion, the distinct cognitive profiles of these subtypes highlight the importance of considering cognitive subtypes in clinical practice and intervention design. Researchers and clinicians with an interest in understanding how functional brain profiles relate to differences in behavior can glean valuable insights from these findings. Tailoring support to specific subtypes may optimize cognitive functioning and inform targeted interventions for individuals with different cognitive profiles.
UPPS The UPPS impulsive behavior scale revealed significant differences among subtypes for negative urgency and positive urgency. In the full sample, negative urgency showed significant differences among subtypes (F(3, 135) = 12.197, p < 0.001), with post-hoc comparisons in Sample 1 and Sample 2 revealing that Subtype 1 had lower negative urgency scores compared to Subtypes 2 and 3, and Subtype 3 had higher scores than Subtype 4 (Robust = True). Positive urgency also showed significant differences across subtypes in the full sample (F(3, 136) = 21.191, p < 0.001), with post-hoc comparisons in Sample 1 and Sample 2 indicating that Subtype 1 had lower scores than Subtypes 2 and 3, and Subtype 3 had higher scores than Subtype 4 (Robust = True). On the other hand, no robust differences were observed for perseverance, premeditation, and sensation-seeking. Although the full sample displayed significant differences for perseverance (F(3, 133) = 8.109, p < 0.001) and premeditation (F(3, 132) = 6.555, p < 0.001), these differences were not consistent across the subsamples. In summary, the findings highlight the importance of considering impulsive behavior subtypes in clinical practice and intervention design, particularly when examining negative urgency and positive urgency profiles. Further research is needed to explore the possible implications of these impulsive behavior differences for understanding and managing impulsivity-related disorders.
NBACK The n-back task analysis revealed significant differences among subtypes for mean reaction time (RT) and total rate correct. In the full sample, mean RT for correct responses displayed significant differences among subtypes (F(3, 103) = 89.368, p < 0.001). Post-hoc comparisons in Sample 1 and Sample 2 showed that Subtype 1 had significantly faster mean RTs compared to all other subtypes, and Subtype 3 had the slowest mean RTs (Robust = True). Similarly, the total rate correct measure also showed significant differences across subtypes in the full sample (F(3, 102) = 93.226, p < 0.001). Post-hoc comparisons in Sample 1 and Sample 2 indicated that Subtype 1 had a significantly higher total rate correct compared to all other subtypes, and Subtype 3 had the lowest rate (Robust = True). On the other hand, no significant differences were observed for the total number correct measure among the subtypes, both in the full sample (F(3, 101) = 0.236, p = 0.872) and in the subsamples (Sample 1: F(3, 0) = 0.783, p = 0.503; Sample 2: F(3, 50) = 0.942, p = 0.419). In summary, the n-back task performance findings highlight the existence of significant differences in mean RT and total rate correct among the subtypes, particularly with Subtype 1 demonstrating faster mean RTs and a higher total rate correct compared to other subtypes. These results may have implications for understanding cognitive functioning differences among various subtypes, which could inform the development of targeted interventions or treatment approaches.
Adversity The analysis of adversity measures revealed significant differences among the subtypes. In the full sample, there were significant differences in adversity scores across subtypes (F(3, 137) = 50.443, p < 0.001). Post-hoc comparisons in both Sample 1 and Sample 2 demonstrated that Subtype 3 had significantly higher adversity scores compared to all other subtypes, while Subtype 4 had the lowest scores (Robust = True). In Sample 1, the differences among subtypes were also significant (F(3, 35) = 34.116, p < 0.001). Similar to the full sample, post-hoc comparisons showed that Subtype 3 had the highest adversity scores, while Subtype 4 had the lowest (Robust = True). In Sample 2, significant differences were again observed among subtypes (F(3, 86) = 17.735, p < 0.001). Post-hoc comparisons revealed that Subtype 3 had significantly higher adversity scores compared to Subtypes 1, 2, and 4, with Subtype 4 having the lowest scores (Robust = True). In summary, these findings indicate that there are significant differences in adversity scores among the subtypes, particularly with Subtype 3 experiencing the highest levels of adversity and Subtype 4 the lowest. This information could be valuable for understanding the relationship between adversity and various subtypes, potentially informing the development of targeted interventions or support strategies to address these disparities.
We observed several key differences across our RSFC subtypes in parental education, household income, marital status, race/ethnicity, and sex at birth. Age did not significantly differ across the subtypes. Parental education revealed distinct patterns among the subtypes. Subtype 3 had a higher proportion of children with parents holding less than a high school diploma and a lower proportion of parents with post-graduate degrees than the other subtypes. In contrast, Subtype 4 showed a higher percentage of children with parents possessing post-graduate degrees and a lower percentage of parents with less than a high school diploma. The household income distribution also varied, with Subtype 3 having the highest proportion of children from households earning less than $50,000 annually and the lowest proportion from households earning more than $100,000 annually. Marital status indicated that Subtype 3 had the highest percentage of children with unmarried parents, while Subtype 4 had the highest percentage of children with married parents. Regarding Race/Ethnicity, Subtype 3 had the highest proportion of Black and Hispanic children and the lowest proportion of White children. Lastly, the sex at birth distribution revealed a slightly higher proportion of males in Subtype 2 than Subtype 3. These demographic differences offer insights into the distinctive characteristics of each RSFC subtype, emphasizing the importance of considering sociodemographic factors in understanding functional connectivity patterns in the human brain.
Bootstrap Reproducibility
The bootstrap analysis provided additional support for the reliability and robustness of the findings on CBCL measures. For instance, in the full sample, aggressive behavior showed significant differences across subtypes (bootstrapped F = 5.767, bootstrapped p = 0.007). This result held in Sample 1 (bootstrapped F = 4.437, bootstrapped p = 0.025), although it was not significant in Sample 2 (bootstrapped F = 2.11, bootstrapped p = 0.272). Attention problems also showed significant differences in the full sample (bootstrapped F = 12.019, bootstrapped p < 0.001), Sample 1 (bootstrapped F = 8.673, bootstrapped p = 0.001), and Sample 2 (bootstrapped F = 4.61, bootstrapped p = 0.039). The externalizing factor also demonstrated significant differences in the full sample (bootstrapped F = 8.836, bootstrapped p < 0.001), Sample 1 (bootstrapped F = 5.675, bootstrapped p = 0.009), and Sample 2 (bootstrapped F = 3.834, bootstrapped p = 0.086), although the latter was marginally significant. Rule-breaking behavior yielded significant differences across subtypes in the full sample (bootstrapped F = 14.303, bootstrapped p < 0.001), Sample 1 (bootstrapped F = 6.901, bootstrapped p = 0.004), and Sample 2 (bootstrapped F = 8.187, bootstrapped p = 0.002). These results reinforce the conclusions drawn from the initial analyses, demonstrating that the differences among subtypes in various CBCL measures are robust and reliable. The bootstrapping approach strengthens the confidence in these findings, allowing for more accurate estimations of the relationships between subtypes and CBCL measures, ultimately helping inform targeted interventions and support strategies.
The bootstrap analysis of cognitive measures further substantiates the robustness and reliability of the findings across subtypes. For example, CardSort_r showed significant differences across subtypes in the full sample (bootstrapped F = 30.511, bootstrapped p < 0.001), Sample 1 (bootstrapped F = 20.231, bootstrapped p < 0.001), and Sample 2 (bootstrapped F = 11.509, bootstrapped p < 0.001). Similarly, LMT_r demonstrated significant differences in the full sample (bootstrapped F = 24.384, bootstrapped p < 0.001), Sample 1 (bootstrapped F = 17.776, bootstrapped p < 0.001), and Sample 2 (bootstrapped F = 8.536, bootstrapped p < 0.001). The Picture_r measure also revealed significant differences across subtypes in the full sample (bootstrapped F = 19.081, bootstrapped p < 0.001), Sample 1 (bootstrapped F = 11.745, bootstrapped p < 0.001), and Sample 2 (bootstrapped F = 8.658, bootstrapped p < 0.001). Lastly, the pc3_new_r measure displayed significant differences in the full sample (bootstrapped F = 35.152, bootstrapped p < 0.001), Sample 1 (bootstrapped F = 23.028, bootstrapped p < 0.001), and Sample 2 (bootstrapped F = 14.182, bootstrapped p < 0.001). These results corroborate the conclusions from the original analysis, highlighting the robust and reliable differences among subtypes in various cognitive measures. The bootstrapping approach bolsters the confidence in these findings, allowing for more precise estimations of the associations between subtypes and cognitive measures. This information is crucial for guiding targeted interventions and support strategies that can enhance cognitive functioning in the different subtypes.
The bootstrap analysis of UPPS impulsivity, adversity, and n-back measures further supports the robustness and reliability of the findings across subtypes. In the UPPS impulsivity measures, negative_urgency showed significant differences across subtypes in the full sample (bootstrapped F = 8.072, bootstrapped p = 0.001), Sample 1 (bootstrapped F = 3.628, bootstrapped p = 0.05), and Sample 2 (bootstrapped F = 4.9, bootstrapped p = 0.026). Similarly, positive_urgency exhibited significant differences in the full sample (bootstrapped F = 13.766, bootstrapped p < 0.001), Sample 1 (bootstrapped F = 8.834, bootstrapped p = 0.001), and Sample 2 (bootstrapped F = 5.785, bootstrapped p = 0.015). In adversity measures, significant differences across subtypes were observed in the full sample (bootstrapped F = 32.254, bootstrapped p < 0.001), Sample 1 (bootstrapped F = 21.828, bootstrapped p < 0.001), and Sample 2 (bootstrapped F = 11.673, bootstrapped p < 0.001). Lastly, for n-back measures, nb_mean_rt_corect displayed significant differences across subtypes in the full sample (bootstrapped F = 56.892, bootstrapped p < 0.001), Sample 1 (bootstrapped F = 31.35, bootstrapped p < 0.001), and Sample 2 (bootstrapped F = 26.19, bootstrapped p < 0.001). Additionally, nb_total_rate_correct revealed significant differences in the full sample (bootstrapped F = 59.32, bootstrapped p < 0.001), Sample 1 (bootstrapped F = 31.112, bootstrapped p < 0.001), and Sample 2 (bootstrapped F = 28.665, bootstrapped p < 0.001). These results confirm the conclusions from the original analysis, emphasizing the robust and reliable differences among subtypes in various impulsivity, adversity, and n-back measures. The bootstrapping approach strengthens the confidence in these findings, providing more accurate estimations of the associations between subtypes and these measures. This information is vital for developing targeted interventions and support strategies that can address impulsivity, adversity, and cognitive functioning in different subtypes.