Brain functional connectivity, but not neuroanatomy, captures the interrelationship between sex and gender in preadolescents

Abstract

Understanding sex differences in the adolescent brain is crucial, as these differences are linked to neurological and psychiatric conditions that vary between males and females. Predicting sex from adolescent brain data may offer valuable insights into how these variations shape neurodevelopment. Recently, attention has shifted toward exploring socially-identified gender, distinct from sex assigned at birth, recognizing its additional explanatory power. This study evaluates whether resting-state functional connectivity (rsFC) or cortical thickness more effectively predicts sex and sex/gender alignment (the congruence between sex and gender) and investigates their interrelationship in preadolescents. Using data from the Adolescent Brain Cognitive Development (ABCD) Study, we employed machine learning to predict both sex (assigned at birth) and sex/gender alignment from rsFC and cortical thickness. rsFC predicted sex with significantly higher accuracy (86%) than cortical thickness (75%) and combining both did not improve the rsFC model's accuracy. Brain regions most effective in predicting sex belonged to association (default mode, dorsal attention, and parietal memory) and visual (visual and medial visual) networks. The rsFC sex classifier trained on sex/gender aligned youth was significantly more effective in classifying unseen youth with sex/gender alignment than in classifying unseen youth with sex/gender unalignment. In females, the degree to which their brains' rsFC matched a sex profile (female or male), was positively associated with the degree of sex/gender alignment. Lastly, neither rsFC nor cortical thickness predicted sex/gender alignment. These findings highlight rsFC's predictive power in capturing the relationship between sex and gender and the complexity of the interplay between sex, gender, and the brain's functional connectivity and neuroanatomy.

Keywords: ABCD; Adolescent Brain Cognitive Development Study; adolescence; brain networks; cortical thickness; gender; machine learning; resting-state functional connectivity; sex.

Figure 1.. Flowchart of the support vector machine (SVM) model construction.

The dataset was split in two groups: Aligned (participants with sex/gender alignment) and unaligned (participants with sex/gender unalignment). The aligned group was split in an aligned training group (80%) and an aligned hold-out testing group (20%). A nested five-fold cross-validation (5F-CV) was employed, with the inner 5F-CV determining the optimal parameter C and the outer 5F-CV estimating the generalizability of the model. The final, optimal model was subsequently applied in the held-out aligned testing group and the unaligned group and model performance was evaluated. The forward slash in the groupings denotes the different possible group splits (e.g. The aligned training group was split in four test/train group pairs with 327/1903 subjects and one test/train group pair with 328/1308 subjects).

Figure 2.. Brain pattern analysis using support vector machine (SVM) learning predicts participant sex based on functional connectivity.

(A) SVMs with nested five-fold cross-validation (5F-CV) were used to construct univariate and multivariate models that classified participants as male or female (sex assigned at birth). These models used resting-state functional connectivity (rsFC), cortical thickness, and rsFC/cortical thickness combined as predictive features. The receiver operating characteristic (ROC) curve of the rsFC resulting model is depicted. The model classified participants as male or female with 86% accuracy. Inset histogram shows distribution of permuted accuracies. The accuracy from real (nonpermuted) data is represented by the dashed red line. (B) McNemar’s tests comparing the rsFC, cortical thickness, and combined rsFC/cortical thickness sex classifiers revealed a statistically significant difference between the rsFC and cortical thickness performance (χ² = 10.96, p < 0.001) and between the combined rsFC/cortical thickness and cortical thickness performance (χ² = 10.96, p < 0.001), but a non-significant difference between the rsFC and combined rsFC/cortical thickness performance (χ² = 0.01, p = 93). (C) To understand which networks contributed the most to the prediction, the parcel-wise feature weights’ root mean square was calculated and then averaged for each network. The most important features in the rsFC model were found in the visual (visual and medial visual), default mode, dorsal attention, and parietal memory networks. (D) The top 10% of cortical parcels in terms of feature importance in the rsFC SVM model. The parcel-wise feature weights’ root mean square was calculated for each parcel. rsFC, resting-state functional connectivity; CT, cortical thickness; NS, not significant; DMN, default mode network; VIS, visual network; MEDVIS, medial visual network; FPN, frontoparietal network; DAN, dorsal attention network; LANG, language network; SAL, salience network; PMN, parietal memory network; AMN, action-mode network; PREMOT, premotor network, SMH, somatomotor hand network; SMM, somatomotor mouth network; SMF, somatomotor foot network; SCAN, somato-cognitive action network; AUD, auditory network; CAN, contextual association network; NONE, subcortical and cerebellar structures.

Figure 3.. Correlations between resting-state functional connectivity (rsFC) support vector machine (SVM) classification scores and sex/gender alignment scores.

(A) A significant positive correlation was observed in females between the rsFC sex classification scores and the sex/gender alignment scores. (B) A significant negative correlation between the rsFC classification and sex/gender alignment scores was observed in males. Higher sex/gender alignment scores (x-axis) indicate greater sex/gender alignment. Higher SVM scores (y-axis) indicate a stronger correspondence to the neural connectivity patterns associated with the sex labeled as either female or male. Results in bold indicate statistical significance.