The data used were collected from participants enrolled in the Human Connectome Project-Aging (HCP-A) study (Bookheimer et al., 2019). Neuropsychological data were drawn from the 2.0 release, which included updates to these data. Imaging data were from the 1.0 release of the HCP-A dataset, as raw imaging data was not updated for the 2.0 release. Imaging data consisted of 689 healthy subjects aged 36 to 100 from four data collection sites. See Bookheimer et al. (2019) for full exclusion criteria. As described previously (Ficek-Tani et al., 2022), we implemented additional exclusion criteria based on motion (see below for details), missing data, and anatomical abnormalities. After exclusion, the remaining sample size was n = 579 (330 female; 249 male).

Participants were well-matched in age, race, ethnicity, years of education, and handedness, but women outnumbered and outperformed men in global cognitive function (Montreal Cognitive Assessment), in-scanner memory task performance (FaceName task), and verbal learning (Rey Auditory Verbal Learning Test; Table 1). Participants self-identified their sex at birth as male or female. While an “Other” option for sex was offered by the HCP-A study, no participants chose this option; gender identity was not assessed.

All subjects enrolled in HCP-A were scanned in a Siemens 3T Prisma scanner with 80 mT/m gradients and 32-channel head coil. In addition to acquiring four resting-state fMRI (rfMRI) and three task-fMRI (tfMRI) scans per subject, structural MRI data [including one T1-weighted (T1w) scan] were also collected (Harms et al., 2018). In this study, we focus on the seven fMRI scans.

A multi-echo MPRAGE sequence [refer to Harms et al. (2018) for scanning parameter details] was used for all T1w scans. A 2D multiband (MB) gradient-recalled echo (GRE) echo-planar imaging (EPI) sequence (MB8, TR/TE = 800/37 ms, flip angle = 52°) was used for all fMRI scans.

For each subject, four rfMRI scans consisting of 488 frames and lasting 6.5 min each (for a total of 26 min) were acquired, during which participants were instructed to remain awake while viewing a small white fixation cross in the center of a black background. The rfMRI scans were split between two sessions that occurred on the same day, with each session including one rfMRI with an anterior to posterior (AP) phase encoding direction and one rfMRI with a posterior to anterior (PA) direction.

The HCP-A includes the following three fMRI tasks, which were all programmed in PsychoPy (Peirce, 2007, 2008) and collected with PA phase encoding direction: Visuomotor (VisMotor), Conditioned Approach Response Inhibition Task (“CARIT” Go/NoGo task), and FaceName (Bookheimer et al., 2019). As below, we focus on the FaceName task scan both because of its relevance to short-term memory performance and because models derived from this scan outperform models derived from other scans. In the FaceName task, three blocks (encoding, distractor, and recall blocks) are repeated twice for each set of faces, totaling to a single, 276-second run. See Harms et al. (2018) for full details on the HCP-A structural and functional MRI imaging parameters, and see Bookheimer et al. (2019) for full details on tfMRI task administration.

The preprocessing approach has been described elsewhere (Greene et al., 2018; Horien et al., 2019). MPRAGE scans were skullstripped with optiBET (Lutkenhoff et al., 2014) and nonlinearly registered to the MNI template in BioImage Suite (BIS; Joshi et al., 2011). BIS was used to linearly register each participant's mean functional scan to their own MPRAGE scan. Skull-stripped and registered data were visually inspected for structural abnormalities and distortions, and participants were excluded for significant structural lesions including meningioma, vascular abnormalities or ventriculomegaly significant enough to distort cortical anatomy, as well as for significant scanner artifacts. Functional data were motion-corrected using SPM8; participants whose scans showed maximum mean frame-to-frame displacement (FFD) above 0.3 mm were excluded to limit motion artifacts (Greene et al., 2018; Horien et al., 2018, 2019; Ju et al., 2020). Using Wilcoxon rank sum tests, we determined no differences in mean FFD between female and male subjects across all seven scan types (Supplementary Table 1). Linear, quadratic, and cubic drift, a 24-parameter model of motion (Satterthwaite et al., 2013), mean cerebrospinal fluid signal, mean white matter signal, and global signal were regressed from the data as described in Ficek-Tani et al. (2022).

Because we were interested in predictors of memory performance, we used performance on the FaceName task and the Rey Auditory Verbal Learning Test (RAVLT) as outcomes for our predictive models. For the FaceName task, participants were shown a total of 10 distinct faces, resulting in a maximum FaceName-Total Recall (FN-TR) score of 10 correctly identified faces. We also assessed both the learning (L) and immediate recall (IR) metrics from the RAVLT (Bean, 2011), a standard neuropsychological measure of declarative memory. In this assessment, a 15-word list is read to the participant, who is then asked to verbally recall as many as possible, five times. The total number of words recalled during this five-trial “learning period” sums to a RAVLT-L (“learning”) score out of 75 words. After being read a separate (interference) list and asked to recall it, the participant is read List A again, and the number of correctly-recalled words in this sixth trial is collected as the RAVLT-IR (“immediate recall”) score. RAVLT-IR is a sensitive metric for early-stage AD (Estévez-González et al., 2003).

To predict memory performance using both rfMRI and tfMRI data from HCP-A, we used connectome-based predictive modeling (CPM), the details of which are described elsewhere (Shen et al., 2017).

In brief, connectivity matrices were constructed from each fMRI scan using the Shen 268-node atlas (Shen et al., 2013). These matrices and the memory performance scores of each participant were used to create our predictive models. Three subject groups were analyzed: all subjects, female-only, and male-only. Edges from connectivity matrices for each subject per scan were correlated to the three aforementioned memory performance measures, totaling to seven connectivity matrices and three memory scores per subject (21 total correlated matrices). Motion and age covariates were also included in the CPM analyses to account for in-scanner head motion, age, and their interaction in our predictions, as previously done (Scheinost et al., 2021; Dufford et al., 2022; Horien et al., 2022).

Using 5-fold cross validation, connectivity matrices and memory scores were divided into independent training (subjects from four of the folds) and testing (subjects in left-out fold) sets. Edge strength and memory were linearly related within the training set, and using a feature selection threshold of p = 0.01, a consensus connectivity matrix including only the edges most strongly positively or negatively correlated to memory was generated. Edge strengths in each subject's connectivity matrix corresponding to the consensus matrix were summed into a single-subject connectivity value. A predictive model built using the linear relationship between the single-subject connectivity values and memory score was applied to the subjects in the testing set to generate memory performance predictions.

For all subject groups, Spearman's correlation and root mean square error [defined as: RMSE (predicted, observed) = √(1/n∑(i=1)n(actuali-predictedi)2)] were used to compare the similarity between predicted and observed memory scores to assess predictive model performance. After performing 1,000 iterations of each CPM analysis, we selected the median-performing model to represent the model's overall performance. To compare model performances between female and male groups for each fMRI scan, we used Wilcoxon rank sum tests.

We also tested our models against randomly permuted models by randomly shuffling participant labels prior to attempting to predict memory scores. After performing 1,000 iterations of this permutation, we calculated the number of times the permuted predictive accuracy was greater than the median unpermuted prediction accuracy to generate a non-parametric p-value, as done in Scheinost et al. (2021):

where #{rho_null ≥ rho_median} indicates the number of permuted predictions numerically greater than or equal to the median of the unpermuted predictions. We applied the Benjamini-Hochberg procedure to these non-parametric p-values to control for multiple comparisons and correct for 21 tests for each of our three subject groups (Benjamini and Hochberg, 1995).

To visualize sex differences at the network level, we first split the aforementioned consensus matrix into two binarized matrices (a “positive” matrix containing edge with significant positive correlations to memory and the other “negative” matrix of edges with significant negative correlations to memory) for each predictive model. Categorization of nodes by functional network was determined using the 10-network parcellation of the Shen 268-node atlas (Horien et al., 2022). In this network grouping, the medial frontal (MF) network also includes some temporal and frontal nodes which often cluster with the DMN. Inter-network edges were defined as the number of significant edges between each pair of networks normalized by the total number of edges between the same network pair. As done in previous work, we defined edges as “significant” if they appear in at least 2 out of 5 folds in 40% of 1,000 iterations of CPM to minimize noise while retaining meaningful connections (Rosenberg et al., 2016; Yip et al., 2019; Horien et al., 2022). In addition to using heatmaps to visualize the inter-network edges of both female and male groups separately, we subtracted male-group positive edges from female-group positive edges (and the same with the negative edges) across corresponding matrix cells to evaluate the inter-network sex differences. Following this, we used the same internally cross-validated procedure as above to apply our sex-based models on subjects of the opposite sex (i.e., testing male-data-trained models on female subject data, and vice versa) and see if they were adequate predictors for the other sex.

Intra-network analyses were performed similarly to inter-network analyses above. Edges from binarized positively and negatively correlated connectivity matrices were summed across the 5 folds and 1,000 iterations to generate a single value for each edge. These values were then used to generate the intra-DMN edge heatmap, with values ranging from −5,000 (maximum negatively correlated) to 5,000 (maximum positively correlated value). To evaluate differences in the “top-performing” nodes according to sex, individual edge values were summed across each row from the matrices, and divided by 2 to account for the symmetric nature of the matrix, generating a summed vector (SV).

We evaluated network segregation, a measure of the relative strength of within-network connections to between-network connections, using a novel association ratio metric. We defined the association ratio as the weighted sum of all edges within the network of interest, normalized by the weighted sum of all edges between this network and the whole set of regions of interest. Higher association ratio is therefore indicative of higher network segregation. To compare network segregation levels between sexes, we calculated and compared (using two-sample t-tests) the association ratio for certain networks of interest in women and men for each scan type. Benjamini-Hochberg correction (see above) was applied to correct for 7 significance tests (for each model) across the 4 networks.

Data from the HCP-A study are openly available (https://www.humanconnectome.org/study/hcp-lifespan-aging/data-releases). Image preprocessing was performed using BioImageSuite, a publicly-available software (https://medicine.yale.edu/bioimaging/suite/). Scripts for running CPM are available through GitHub (https://github.com/YaleMRRC/CPM). Other MATLAB scripts for CPM analyses can be found at https://github.com/frederickslab/CPM_HCP-A_sex_difference_study. Custom MATLAB colormap palettes were derived from ColorBrewer (http://colorbrewer.org/; Brewer, 2022).

Please see Supplementary Methods/Results for details on model comparisons, including comparisons between models derived separately for each sex. Briefly, we trained and cross-validated models using functional connectivity data from all 7 scans to predict memory performance scores. Whole-group models robustly predicted all memory measures, with accuracies ranging from Spearman's rho = 0.21 (RMSE = 3.34, p < 0.0001) to rho = 0.45 (RMSE = 2.67, p < 0.0001) across all models (Supplementary Figure 1). Models using the FaceName tfMRI scan consistently outperformed all other models; we therefore proceeded with models from this scan for the remaining analyses.

Visualizations of inter-network edges (number of significant edges normalized by network size) across all FaceName tfMRI models revealed differences in key edges predicting memory score for each sex. In particular, edges within the DMN and visual (visual I [VI], visual II [VII], and visual association areas [VAs]) networks showed the largest differences (Figure 1, Supplementary Figure 5). Given previous work showing measures of declarative verbal memory (including RAVLT metrics) can be predicted from the gray matter density of DMN structures, and because lower RAVLT-IR scores are associated with preclinical AD, we concentrated on the RAVLT-IR predictors derived from FaceName tfMRI models (Estévez-González et al., 2003; Moradi et al., 2017). In addition to visualizing the inter-network edges of females and males separately, we subtracted male-group edges from female-group edges across corresponding heatmap cells to evaluate inter-network differences between the sexes (Figure 1).

Both sexes show positive predictors with intra-DMN edges, with female scores predicting intra-DMN connectivity more strongly than those of males. Female positive predictors also relied more strongly on intra-VI edges than those of males, while male positive predictors relied more strongly on the intra- and inter-network connectivity of the VII and VAs networks relative to those of females. Both sexes displayed negative predictors with edges between DMN and visual networks; however, males show more negative predictors with edges between the MF and VII networks, as well as between the DMN and VII networks, relative to females. Additionally, testing our sex-based models on subjects of the opposite sex revealed that our female models successfully predicted outcomes for male subjects across all our memory tasks, while our male models successfully predicted only the RAVLT-L outcome for female subjects (see Supplementary Table 10 for model performance statistics).

Given the preferential contribution of intra-DMN edges to the female models, we examined all intra-DMN edges and evaluated their strengths in male and female models. To do so, we generated a heatmap of intra-DMN edges (Figure 2). In the RAVLT-IR model, we found that edges from more posterior DMN nodes were preferentially increased in females as opposed to males. This trend held true for the RAVLT-L model and FN-TR models (Supplementary Figure 7). Negatively correlated edges negligibly contributed to both male and female models (Figure 2, Supplementary Figure 7). Both sexes displayed strong connections in the left posterior cingulate cortex (L PCC) and precuneus, known hubs of the DMN.

To summarize node-level differences, we summed the number of edges associated with each node and found consistent female preference for activity of the right posterior inferior parietal lobe (R pIPL) and left anterior medial prefrontal cortex (L amPFC)/paracingulate cortex (Figure 2). The R pIPL was consistently and preferentially elevated in all female models analyzed (Supplementary Figure 7).

We then evaluated and compared a metric of network segregation (see Methods, “Network Segregation Analysis”) within the DMN and visual (VI, VII, VAs) networks between females and males, given the strong brain-behavior correlations in these networks across all memory performance outcomes. Our analysis demonstrated increased network segregation of the DMN in females relative to males, and increased network segregation of VII and VAs in males relative to females (Table 2). Additionally, these findings echoed our previous CPM analysis results in that we also observed sex differences in neurobiological organization.