Prevalent and sex-biased breathing patterns modify functional connectivity MRI in young adults

Resting-state fMRI (functional connectivity MRI) measures spontaneous correlations in BOLD signals to infer functional organization of the human brain. Breathing modulates arterial CO2, cerebral blood flow, and hence fMRI signals, yet the breathing characteristics of healthy young adults during rest in MRI scanners are poorly characterized despite their potential to systematically bias fMRI. The authors examined respiration and fMRI jointly in the Human Connectome Project (HCP) Young Adult dataset to identify prevalent respiratory patterns during rest that influence fMRI signals and connectivity. They report two common deviations from normal tidal breathing: single deep breaths and a previously undescribed burst pattern marked by serial tapering of breathing depth. The study asks how these patterns affect fMRI signals and covariance, whether they differ in timescale and cardiovascular correlates, and whether their prevalence differs by sex.

Prior work shows respiration-related fMRI fluctuations can be modeled using measures such as respiratory variation (RV), respiratory volume per time (RVT), and related envelopes (e.g., RETROICOR and respiration response function approaches). Various forms of disordered breathing (e.g., periodic, cluster, ataxic) are known in clinical contexts and often linked to heart disease or neurological injury; chemoreflex mechanisms governing pCO2 set points and apneic thresholds influence respiratory stability. Previous fMRI respiration studies typically involved small samples and short recordings. Literature on sex differences in sleep-related breathing disorders and chemoreflex physiology suggests men tend to exhibit more central apnea/periodic breathing, influenced by sex hormones (e.g., testosterone elevating apneic threshold and reducing CO2 reserve). Studies of arousal/sleep, global fMRI signals, and sensorimotor network prominence also provide context for interpreting respiration-related global and spatially specific effects.

Dataset: HCP Young Adult release. Inclusion: 440 healthy young adults (ages 22–36, mean 28.6; 228 males, 212 females) with four 14.4-min resting-state scans each (two per day across two days) and high-quality concurrent physiology (abdominal respiratory belt and finger pulse oximeter at 400 Hz). Total analyzed rest time: 440 hours. Preprocessing and visualization: Minimally preprocessed and FIX-ICA denoised fMRI data were used. Masks for gray matter, white matter, ventricles, cerebellum, and subcortical nuclei were derived; within-mask 6 mm smoothing aided gray plot visualization. Gray plots (all in-brain voxel time series arranged by tissue compartment) were generated for all 1,760 scans to visually identify respiration-related patterns. Respiratory signal processing: Respiratory belt traces were filtered to remove spikes and smoothed; derived measures included ENV (envelope over 10 s), RV (6 s moving SD), and RVT (peak-to-trough amplitude divided by inter-peak interval). Pulse oximetry peaks yielded heart rate. Head motion (framewise displacement, FD) was computed from realignment parameters with respiratory band-stop filtering (0.2–0.5 Hz). DVARS was computed as standard data quality metric. Event-related analysis: Visual identification of 35 burst onsets, 35 deep-breath onsets, and 35 non-respiratory motion onsets; random onsets served as control. For each event, data from -30 to +60 s were extracted. Heatmaps and mean trajectories were compared to random events using two-sample t-tests (p<0.001 thresholding displayed). Global fMRI signal dynamics, motion, DVARS, and heart rate responses were characterized. Network-level spatiotemporal effects around events were analyzed using the 333-parcel Gordon atlas; correlations were computed from -10 to +40 s windows around onsets. Significance masking used 10,000 permutation tests (p<0.05). Rater scoring and inter-rater reliability: Two trained raters independently scored scans (N=1,596 scans from 399 subjects) for presence/absence of deep breaths and bursts; pattern scores (0–4) were sums across four scans. Cohen’s kappa assessed reliability; chi-squared tests assessed sex differences. Automated indexing: An algorithm combined respiratory-belt-derived priors for each pattern with template matches of global fMRI signals to yield pattern indices per scan, down-weighted by variability in respiratory rate. Indices were compared with rater scores and examined for sex differences. Group formation for contrasts: Three groups (N=21 unrelated subjects each) selected based on scan-level breathing characteristics: burst group (pronounced bursts, few deep breaths), deep-breath group (pronounced deep breaths, few bursts), and clean group (neither). Between-group contrasts used permutation testing. Global functional connectivity (gFC): Median pairwise correlation among all gray matter voxels per scan (Fisher z) averaged across scans per subject. Associations with pattern scores/indices were tested via Pearson correlations and multiple linear regression; ANCOVA modeled gFC as a function of sex, head size (intracranial volume), and respiratory variables, with motion and DVARS as covariates in additional models. Whole-scan connectivity analyses: Correlation matrices for entire scans contrasted between groups, within subjects (scans with vs without patterns), and via regression betas across subjects, including multiple preprocessing strategies: minimally preprocessed (MP), FIX-ICA denoised (Post-FIX), MP+global signal regression (GSR), MP+motion regression+censoring (DVARS>2), and combinations thereof. Significance used 10,000 permutations, p<0.05.

• Two prevalent respiratory patterns beyond eupnea were identified: (1) single deep breaths and (2) bursts (serial, rhythmic tapering in breathing depth over minutes). Both produce pan-brain BOLD modulations but differ in timescale, shape, and spatial effects.
• Event-related BOLD dynamics: Deep breaths show a brief global signal increase followed by a marked decrease (nadir ~15 s, resolution ~30 s). Bursts show slower trajectories with higher peaks and prolonged troughs (nadir >20 s, resolution ~40 s).
• Motion control: Non-respiratory head motions and random onsets did not elicit global fMRI signal changes, indicating respiration—not motion—drives the observed global signals.
• Cardiovascular correlates: Heart rate reliably increased for several seconds after deep breaths; bursts showed no consistent average heart rate change across subjects.
• Sex bias: Bursts were significantly more common in males; deep breaths showed no sex bias. Across scans, bursts were identified in 45% of male scans vs 35% of female scans; deep breaths in 54% of male vs 52% of female scans. Chi-squared tests for bursts were highly significant for both raters (p=3.4e-8 and 8.8e-7). In exploratory groups (each N=21), sex distributions were unbalanced: clean 6/21 males, deep-breath 5/21 males, burst 14/21 males (joint probability p=3.3e-5).
• Rater reliability and algorithm validation: Inter-rater agreement was high (Cohen’s kappa ~0.73–0.78). Pattern scores correlated strongly between raters (r=0.86 for bursts; r=0.90 for deep breaths). Automated indices correlated with rater scores and recapitulated the sex difference for bursts but not deep breaths.
• Global functional connectivity (gFC): gFC increased with pattern prevalence, more strongly for bursts. With rater scores, gFC correlations were r=0.53 (bursts, p≤e-20) and r=0.18 (deep breaths, p=4e-4). With algorithmic indices, r=0.59 (bursts, p≤e-20) and r=0.29 (deep breaths, p=7e-9). Multiple regression showed larger betas for bursts; fits did not differ by sex. Males had higher gFC on average, but ANCOVA showed sex effects became non-significant when accounting for head size and respiratory variables.
• Spatial specificity: Bursts robustly elevated correlations in primary sensorimotor distributions (visual, auditory, somatomotor), whereas deep breaths produced broader elevations that conspicuously avoided these sensorimotor regions. These distinctions persisted across preprocessing strategies; global signal regression altered but did not eliminate significant burst-related effects, indicating lagged non-global structure.
• Temporal occurrence within scans: Both patterns became more likely as scans progressed; increases from minutes 1–4 to 11–14 were significant for bursts in both sexes (p<1e−16) and for deep breaths (p<0.02).
• Sleep/arousal and BMI: Deep-breath group members were more often listed as sleepy by technicians (71% vs 38–41% in other groups), consistent with links to yawning/sleepiness. Bursts were not associated with BMI in either sex, arguing against obstructive apnea as the primary cause in this cohort.

The study demonstrates that common, distinct respiratory patterns during resting-state fMRI—single deep breaths and bursts—systematically modulate global and spatially specific functional connectivity. Bursts, more prevalent in males, have especially strong effects on global functional connectivity and selectively elevate correlations in sensorimotor networks, whereas deep breaths yield pan-brain effects with relative sparing of sensorimotor regions. Because global and spatially structured respiration-induced variance can inflate functional connectivity, sex differences in resting-state connectivity can arise indirectly via sex-biased respiration; indeed, male–female gFC differences disappeared after accounting for head size and respiration. The distinct temporal dynamics, spatial profiles, and cardiac responses suggest different underlying cardio-pulmonary and neurophysiological mechanisms, with bursts plausibly linked to chemoreflex-controlled periodic breathing. A chemoreflex-based account, influenced by sex hormones (testosterone raising apneic threshold and reducing CO2 reserve), may explain male bias in bursts. Findings align with broader literature on global signals, arousal/sleep, and sensorimotor network prominence. The work underscores the importance of recording and modeling respiration in resting-state studies and motivates evaluation of preprocessing/denoising strategies (e.g., GSR, ICA-based denoising) with respect to respiration-related effects.

This work identifies two prevalent respiration patterns in healthy young adults at rest—single deep breaths and male-biased bursts—that differentially and substantially influence resting-state fMRI signals and connectivity, especially global functional connectivity and sensorimotor network correlations. Accounting for head size and these respiratory patterns removes apparent sex differences in gFC. The authors provide human- and algorithm-based detection approaches and show robust spatiotemporal signatures across preprocessing pipelines. Future directions include: disentangling pCO2-driven cerebrovascular effects from neural activity during each pattern; extending analyses across lifespan and clinical/psychiatric populations; testing modulation by sex hormones, arousal/vigilance, and pharmacologic agents (e.g., opioids, sedatives); and standardizing concurrent physiology acquisition and modeling in fMRI pipelines.

• Lack of direct measures of arousal/sleep (no polysomnography or definitive wakefulness markers); reliance on technician sleepiness notes limits inference about vigilance effects.
• Cardiac measures were limited and required substantial manual verification; not all subjects had usable cardiac data.
• Automated indexing can be challenged by disorganized or haphazard breathing, necessitating discounting scans with highly variable respiratory rates.
• Although BMI was not associated with bursts, some burst-like patterns could still reflect obstructive events in individual cases; central vs obstructive contributions cannot be definitively separated here.
• Group identities tied to restricted HCP variables (psychiatric/substance use) limit full transparency; generalizability beyond healthy young adults requires further study.
• While motion was carefully evaluated and did not produce global signals, residual confounds or unmeasured variables could contribute to observed associations.