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 they relate to sex-specific neurological and psychiatric conditions. Predicting sex from adolescent brain data may reveal how these differences influence neurodevelopment. Recently, attention has shifted toward socially-identified gender (distinct from sex assigned at birth) recognizing its explanatory power. This study evaluates whether resting-state functional connectivity (rsFC), cortical thickness, or cortical volume better predicts sex and sex/gender alignment (congruence between sex and gender) in preadolescents. Using Adolescent Brain Cognitive Development (ABCD) Study data and machine learning, rsFC predicted sex more accurately (85 %) than cortical thickness (76 %) and cortical volume (70 %). Brain regions most predictive of sex belonged to association (default mode, dorsal attention, parietal memory) and visual networks. The rsFC classifier trained on sex/gender aligned youth classified more accurately unseen youth with sex/gender alignment (n = 2013) than unalignment (n = 1116). The female rsFC sex profile was positively associated with sex/gender alignment, while in males, there was a negative association. However, neither brain modality predicted sex/gender alignment. These findings suggest that while rsFC predicts sex in the adolescent brain more accurately, it does not directly capture sex/gender alignment, underscoring the need for further investigation into the neural underpinnings of gender.

Keywords: Adolescence; Adolescent brain cognitive development study; Brain networks; Cortical thickness; Gender; Resting-state functional connectivity; Sex.

Graphical abstract

Fig. 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 performance estimation loop training group was split in three train/test group pairs with 1030/258 subjects and two train/test group pairs with 1031/257 subjects respectively).

Fig. 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, cortical volume, rsFC/cortical thickness, and rsFC/cortical volume 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 85 % 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, cortical volume, and combined rsFC/cortical thickness and rsFC/cortical volume sex classifiers revealed a statistically significant difference in performance between rsFC and cortical thickness (χ² = 7.36, p < 0.01) and between rsFC and cortical volume (χ² = 19.67, p < 0.001). However, the performance differences between cortical thickness and cortical volume, rsFC and rsFC/cortical thickness, and rsFC and rsFC/cortical volume sex classifiers were not significant. (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; HC, hippocampus; AMG, amygdala; BG, basal ganglia; THAL, thalamus; CERB, cerebellum.

Fig. 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.