Predicting Sex from Resting-State fMRI Across Multiple Independent Acquired Datasets

Abstract

Sex is an important biological variable often used in analyzing and describing the functional organization of the brain during cognitive and behavioral tasks. Several prior studies have shown that blood-oxygen-level-dependent (BOLD) functional MRI (fMRI) functional connectivity (FC) can be used to differentiate sex among individuals. Herein, we demonstrate that sex can be further classified with high accuracy using the intrinsic BOLD signal fluctuations from resting-state fMRI (rs-fMRI). We adopted the amplitude of low-frequency fluctuation (ALFF), and the fraction of ALFF (fALFF) features from the automated anatomical atlas (AAL) and Power's functional atlas as an input to different machine learning (ML) methods. Using datasets from five independently acquired subject cohorts and with eight fMRI scanning sessions, we comprehensively assessed unbiased performance using nested-cross validation for within-sample and across sample accuracies. The results demonstrated high prediction accuracies for the Human Connectome Project (HCP) dataset (area under cure (AUC) > 0.89). The yielded accuracies suggest that sex difference is embodied and well-pronounced in the low-frequency BOLD signal fluctuation. The performance degrades with the heterogeneity of the cohort and suggests that other factors,.e.g. psychiatric disorders and demographics influences the BOLD signal and may interact with the classification of sex. In addition, the results revealed high learning generalizability with the HCP scan, but not across different datasets. The intraclass correlation coefficient (ICC) across HCP scans showed moderate-to-good reliability based on atlas selection (ICC = 0.65 [0.63-0.67] and ICC= 0.78 [0.76-0.80].). We also assessed the effect of scan duration on the predictability of sex and showed that sex differences could be detected even with a short rs-fMRI scan (e.g., 2 minutes). Moreover, we provided statistical maps of the brain regions differentially recruited by or predicting sex using Shapely values and determined an overlap with previous reports of brain response due to sex differences. Altogether, our analysis suggests that sex differences are well-pronounced in rs-fMRI and should be considered seriously in any study design, analysis, or interpretation.

Comprehensive evaluation of harmonization on functional brain imaging for multisite data-fusion

“…Among such results, SMA-harmonized detected higher activity of posterior cingulate cortex of females for both ALFF and fALFF, while ComBat-harmonized was unable to transport that discovery from CoRR to FCP. Significantly higher fALFF in males was found in bilateral insula, sensory and motor related cortices, consistent with prior studies (Al Zoubi et al, 2020; Allen et al, 2011). In addition, areas with significant fALFF were highly spatially similar to those reported by Biswal et al (Biswal et al, 2010), showing high replicability and statistical power after SMA harmonization.…”

Because the Machine Can Discriminate: How Machine Learning Serves and Transforms Biological Explanations of Human Difference

“…Indeed, others have argued that the lack of relationship between sample size and number of sex differences 'discovered' is evidence of reporting bias in this literature (David et al 2018). Widespread design flaws in this research have also led to inflated accuracies, including failure to properly control for head size and other confounds (Dhamala et al 2020;Eliot et al 2021;Linn et al 2016;Sanchis-Segura et al 2020), failure to test on large enough and independent samples (Glocker et al 2019;Huf et al 2014;Varoquaux 2018;Zoubi et al 2020), and failure to think through model assumptions (Bzdok 2017;Carlson et al 2018;Schulz et al 2020).…”

Because the machine can discriminate: How machine learning serves and transforms biological explanations of human difference

“…Indeed, others have argued that the lack of relationship between sample size and number of sex differences 'discovered' is evidence of reporting bias in this literature (David et al, 2018). Widespread design flaws in this research have also led to inflated accuracies, including failure to properly control for head size and other confounds (Dhamala et al, 2020;Eliot et al, 2021;Linn et al, 2016;Sanchis-Segura et al, 2020), failure to test on large enough and independent samples (Glocker et al, 2019;Huf et al, 2014;Varoquaux, 2018;Zoubi et al, 2020), and failure to think through model assumptions (Figure 3) (Bzdok, 2017;Carlson et al, 2018;Schulz et al, 2020).…”