Cerebral Oxygen Budgeting: Network-Level BOLD Dynamics During Acute Hypoxia

Abstract

Hypoxia constrains cerebral oxygen availability and challenges brain function and stability. Although hypoxia-responsive functional connectivity (HR-FC) reorganizes rapidly with declining arterial oxygen partial pressure, its relationship to local neurovascular activity remains unclear. We examined time-resolved amplitude of low-frequency fluctuations (ALFF) in blood-oxygenation-level dependent (BOLD) fMRI during graded acute hypoxia in healthy adults, performing a continuous cognitive test (Go/No-go task) with concurrent physiological monitoring. Dynamic ALFF and functional connectivity were estimated using a sliding-window approach and analyzed across large-scale brain networks defined by Schaefer’s 17-network parcellation. Severe hypoxia elicited temporally dissociated responses across modalities. Functional connectivity increased monotonically, whereas ALFF exhibited pronounced nonlinear modulation, including phase-dependent divergence across networks. During hypoxic decompensation, the default mode network (DefaultA) showed marked ALFF suppression, whereas a ventral secondary somatosensory-dominant network (SomMotB) exhibited preferential preservation despite similar engagement in HR-FC. Together, these findings indicate that network-level ALFF captures a distinct yet complementary layer of functional dynamics, with a temporal profile distinct from functional connectivity. Spontaneous BOLD dynamics during acute hypoxia reflect structured network-level modulation rather than a uniform suppression attributable solely to reduced oxygen availability. These findings support a conceptual framework of cerebral oxygen budgeting, in which metabolic constraints reshape functional dynamics across brain networks.

2. Introduction

Oxygen serves as the brain’s metabolic budget ¹, and even brief reductions in its availability can trigger a wide array of physiological and cognitive alterations ². Physiologically, a decline in arterial oxygen partial pressure (P_aO₂) is the earliest and most sensitive indicator of hypoxic exposure, although cerebral oxygen delivery ultimately depends on arterial oxygen content (CaO₂) and blood flow ³. In response to reduced oxygen availability, the brain attempts to maintain metabolic homeostasis by increasing cerebral blood flow through vasodilation and by adjusting the rate of oxygen metabolism to sustain or prioritize energy production ^4–8. Although compensatory, these responses may not fully mitigate the effects of hypoxia, which can lead to temporary or permanent cognitive impairment and neural dysfunction ^9–15. These effects often emerge gradually or sequentially, with timing and severity depending on the depth and duration of hypoxic exposure. Despite extensive observations across physiological, behavioral, and neuroimaging domains, a mechanistic framework that unifies these phenomena remains lacking, motivating a fundamental question, “How does the brain budget its oxygen under hypoxic stress?”

Recent work has shown that acute hypoxia-induced changes in functional connectivity (FC) correlate with reductions in end-tidal oxygen partial pressure (P_ETO₂), a surrogate for arterial oxygen tension (P_aO₂), and emerge at a reproducible physiological threshold ¹⁶. These findings revealed a proactive and structured reconfiguration of large-scale brain networks in response to hypoxic stress, characterized by a selective increase in inter-network connectivity centered on the default mode network (DMN) during hypoxia. Notably, although FC increased following threshold crossing of P_ETO₂, these changes did not track the timing or magnitude of behavioral deterioration. FC reorganization was observed in both severe and mild hypoxia (fraction of inspired oxygen F_iO₂= 7.7% and 11.8%, respectively) ¹⁶. In severe hypoxia, these changes preceded behavioral decline ¹⁶, whereas the mild hypoxia did not elicit significant behavioral deterioration in prior work using the same paradigm ¹⁴. These discrepancies indicate that the functional relevance of FC reconfiguration, particularly how it supports or reflects the brain’s response to hypoxic stress, remains unresolved, as FC does not directly reflect neural or metabolic processes.

Amplitude of low-frequency fluctuations (ALFF), as voxel-wise metrics derived from resting-state blood-oxygenation-level dependent (BOLD) signals, have been widely used to characterize regional brain dysfunction across neurological and psychiatric disorders. ¹⁷. ALFF quantifies the power of BOLD signal fluctuations at low frequencies (0.01–0.10 Hz) and is considered a proxy for regional spontaneous neurovascular activity, whereas FC quantifies the temporal coupling of the BOLD fluctuations between spatially distributed brain regions. Together, these measures capture complementary aspects of BOLD dynamics, reflecting regional signal amplitude and large-scale network coordination. Analogous to dynamic analyses of FC (dFC) that characterize time-varying network organization, time-resolved approaches for ALFF may offer additional insight into rapid physiological transitions, such as those induced by acute hypoxia. Acute hypoxia alters not only neurovascular activity but also physiological rhythms, including cardiac and respiration dynamics. In addition, we examined dynamic amplitudes of high-frequency fluctuations (dAHFF, > 0.10 Hz). Because higher-frequency BOLD components are more sensitive to cardiac, respiratory, and vascular dynamics, dAHFF serves as a complementary measure to dALFF, enabling characterization of hypoxia-related physiological modulation beyond the low-frequency band.

Building on our previous observation that hypoxia-responsive FC (HR-FC) emerges with declining P_ETO₂ and reorganizes large-scale brain networks ¹⁶, the present study investigates the regional and temporal dynamics of ALFF during acute hypoxia and tests the hypothesis that ALFF changes are systematically linked to FC reorganization at the large-scale network level. Although ALFF is traditionally treated as a regional metric, summarizing it at the network level allows its amplitude dynamics to be evaluated alongside FC patterns. Furthermore, by comparing the temporal evolution of P_ETO₂, peripheral oxygen saturation (S_pO₂), behavioral outputs, dFC, dALFF, and dAHFF, we delineate data-informed hypoxic and recovery phases and characterize the phase-dependent response patterns and relative phase alignment. Using this phase-resolved, network-integrated approach, we aim to clarify how the brain coordinates large-scale functional dynamics under acute metabolic stress.

3. Methods and materials

3.1. Participants and acute hypoxia paradigm

Eleven healthy adults (five females / six males; age 26.5 ± 4.5 years, mean ± SD) were recruited for this study. All participants were self-reported as healthy, with no known vascular, respiratory, cardiac, or neurological conditions. Prior to the experimental session, participants were familiarized with the respiratory circuit and the cognitive task outside the MRI suite. Baseline physiological characteristics of the cohort are summarized in Supplementary Table 1. The study was conducted under an Institutional Review Board–approved protocol (Mayo Clinic IRB#16–004354), and all participants provided written informed consent prior to participation. Experimental procedures were performed in accordance with the ethical standards set by the Declaration of Helsinki.

Each participant underwent three 10-min fMRI scans in the following order: normoxia, severe hypoxia, and mild hypoxia. The severe hypoxia scan followed a block design consisting of three minutes of normoxia (F_iO₂ = 21%) (baseline), followed by three minutes of hypoxic gas (F_iO₂ = 7.7%) administration, and four minutes of recovery with normoxic air. In contrast, the normoxia and mild hypoxia scans were performed under continuous delivery of the assigned gas mixture throughout the 10-min acquisition (F_iO₂ = 21% and 11.8%, respectively). A minimum interval of 20 min was maintained between scans. Participants were instrumented with a two-way non-rebreathing valve (2700 Series, Hans Rudolph, Shawnee, KS, USA) attached to a facemask (7450 Series V2, Hans Rudolph, Shawnee, KS, USA). During the normoxic inspirate condition, participants inspired compressed room air (~21% oxygen balanced with nitrogen). Normobaric hypoxia inspirates consisted of pre-mixed tanks containing 11.8% O₂ (moderate hypoxia) and 7.7% O₂ (severe hypoxia) balanced with nitrogen. All gases were delivered from a Douglas bag reservoir containing the pre-mixed gas. During all fMRI scans, participants performed a computer-based Go/No-Go cognitive task ¹⁴.

To ensure subject safety, a board-certified anesthesiologist was present throughout all MRI sessions, including hypoxic challenges, continuously monitoring participant condition and cardiorespiratory parameters.

3.2. Data acquisition

All participants were scanned on a compact 3T MRI system (C3T) ¹⁸, using a 32-channel head coil (Nova Medical, Wilmington, MA, USA).

Functional images were acquired using a gradient-echo echo-planar imaging (GRE-EPI) sequence during each 10-min scan (TR = 2 s, TE = 30 ms, flip angle = 90°, isotropic resolution = 2.5 mm, simultaneous multi-slice acceleration factor = 3, no in-plane acceleration, 300 volumes). A high-resolution T1-weighted anatomical image was acquired for spatial reference using a magnetization-prepared rapid acquisition with gradient echo (MPRAGE) sequence (TR = 5.4 ms, TE = 2.4 ms, TI = 1000 ms, flip angle = 8°, isotropic resolution = 1.0 mm).

Physiological signals were continuously recorded during functional imaging. Arterial blood pressure was monitored noninvasively using finger photoplethysmography (Nexfin, TPD Biomedical Instrumentation). Heart rate, peripheral oxygen saturation (S_pO₂), and breath-by-breath end-tidal O₂ and CO₂ were continuously monitored (Cardiocap/5, Datex-Ohmeda). Participants breathed through a low-dead-space, two-way non-rebreathing valve circuit, with expiratory flow measured using a turbine system (Universal Ventilation Meter, Ventura, CA). Analog signals were collected and converted to digital signals at a rate of ~1,000 Hz using a data acquisition system (PowerLab 16/30 and LabChart 8, ADInstruments, Colorado Springs, CO, USA).

3.3. Cognitive and behavioral measures

During the functional imaging sessions, the participants performed a computer-based Go/No-Go cognitive task, implemented using MATLAB software (R2019b, MathWorks, Natick, MA) ¹⁴. Task stimuli were presented visually using an LCD display system (Nordic NeuroLab, Bergen, Norway). The subjects were instructed to press a button in response to a “Go” stimulus but refrain from pressing the button for a “No-Go” stimulus.

Behavioral performance was quantified using commission error rate, omission error rate, and reaction time. Commission errors were defined as incorrect responses to No-Go stimuli, whereas omission errors reflected failures to respond to Go stimuli. Reaction time was defined as the interval between stimulus presentation and button press. These metrics were continuously recorded throughout the task. To ensure adequate task familiarization, participants completed a 10-min practice session prior to scanning, as previously recommended ¹⁹. Based on prior evidence that commission error rates increase under acute hypoxia and sensitively reflect impaired inhibitory control, commission error rate was selected as the primary behavioral marker of hypoxia-induced cognitive deterioration ¹⁴.

3.4. Data preprocessing and signal derivation

3.4.1. Preprocessing for MR images

All functional MRI data, consisting of 300 consecutive EPI volumes per scan, were preprocessed. Preprocessing steps included removal of the first 10 volumes, de-spiking, slice-timing correction, physiological noise correction using RETROICOR ^20,21, EPI–MPRAGE alignment, inter-volume motion correction, and spatial smoothing with a Gaussian kernel (4 mm full-width at half-maximum, FWHM). Signal intensities were then scaled to a mean value of 100.

Subsequently, nuisance regression was performed using linear regressors including Legendre polynomials (up to 4th order), rigid-body motion parameters, and white matter and cerebrospinal fluid (WM/CSF) signals derived using ANATICOR and CompCor approaches ^22,23. Residual BOLD time series were obtained following regression using the AFNI software package ²⁴. No motion censoring was applied.

T1-weighted anatomical images were processed to segment brain structures using FreeSurfer ²⁵. The resulting anatomical segmentations were used for coregistration to individual EPI images and for projection onto standard cortical surfaces provided by AFNI and SUMA ²⁴.

3.4.2. Pre-defined ROI with intrinsic brain networks

Regions of interest (ROIs) were defined using the 400-region cortical parcellation from the Schaefer atlas ²⁶. The atlas was implemented on the AFNI standard surface model (https://afni.nimh.nih.gov/pub/dist/atlases/SchaeferYeo/) and subsequently transformed into individual subjects’ native volumetric space for analysis of EPI data. Each ROI was assigned to one of 17 intrinsic connectivity networks, enabling analyses at both the ROI and network levels for functional connectivity and spontaneous BOLD signal fluctuations.

3.4.3. Dynamic changes by sliding window

To examine dynamic ALFF (dALFF), mean BOLD time series were extracted from each of the 400 ROIs using the preprocessed residual BOLD signals. ALFF was computed for each ROI using a fast Fourier transform (FFT function) by calculating the square root of the power spectrum within the low-frequency band (0.01 ≤ f ≤ 0.10 Hz). dALFF was estimated using a causal sliding window of 90 seconds with a step size of 2 seconds, resulting in 246 overlapping windows per scan. This procedure generated a dALFF time series for each ROI in each subject. In addition, dAHFF was computed within the frequency band of 0.10 < f ≤ 0.25 Hz using the same FFT-based and sliding-window procedures, with the upper bound constrained by the Nyquist frequency determined by the repetition time (TR = 2 s). dALFF and dAHFF were quantified as relative changes (i.e., percent changes) with respect to a region-specific temporal baseline. A 90-s sliding window was used in the main analysis as a compromise between temporal sensitivity and stability. Similar ALFF response patterns were observed when using alternative window lengths (60 s and 120 s), indicating that the results were robust to window size (Supplementary Fig. 1).

dFC was computed from the residual BOLD signals within the low-frequency band (0.01 ≤ f ≤ 0.10 Hz). For each sliding window, Fisher z-transformed Pearson correlation coefficients were calculated between all ROI pairs, yielding a 400 × 400 functional connectivity matrix (79,800 unique connections) per window and a time-resolved FC series for each subject.

Bulk BOLD signal changes were evaluated using motion-corrected BOLD signals averaged across voxels within the cerebral gray matter.

3.5. Network-level and statistical analyses

3.5.1. Temporal alignment and normalization of multimodal signals

To facilitate direct comparison of temporal trajectories across physiological, neuroimaging, and behavioral measures under acute severe hypoxia, aggregate time series of P_ETO₂, FC, ALFF, and behavioral error rates were z-score normalized and lightly temporally smoothed using a 5-point moving average filter (10 s window given the 2-s sampling interval), implemented with the movmean function in MATLAB. For each metric, baseline values were defined using the pre-hypoxia period (t < 180 s), and all time courses were aligned such that the baseline mean was set to zero. Temporal smoothing was applied to reduce high-frequency noise while preserving low-frequency trends relevant to hypoxia-induced dynamics. This procedure facilitated visualization and comparison of the relative timing and progression of physiological, imaging-derived, and behavioral responses to hypoxia.

3.5.2. Principal component analysis

To identify dALFF patterns associated with hypoxia, we applied principal component analysis (PCA) to the fMRI session that followed a block-design paradigm (normoxia–severe hypoxia–normoxia). This was performed because its structured temporal profile facilitates extraction of hypoxia-related signal patterns.

Prior to PCA, dALFF time series were z-scored within each ROI for each subject to remove subject-specific mean and variance differences. PCA was performed in MATLAB (R2019b) using default settings. The input matrix consisted of 2,706 observations × 400 ROIs, constructed by concatenating dALFF time points from 246 causal sliding windows across 11 subjects (246 × 11 = 2,706). In this formulation, each time point (window) represented one observation and each ROI was treated as a variable. Thus, PCA identified spatial patterns of ROIs whose dALFF values covaried across time and subjects. Components were ranked by the proportion of explained variance, and the dominant components were carried forward for subsequent analyses. For visualization of temporal dynamics, component time courses were then reshaped into a 246 (windows) × 11 (subjects) matrix to enable visualization and comparison of time courses across individuals and phases.

A similar PCA framework was applied to dFC, following our previous work ¹⁶. For this analysis, the input matrix consisted of 2,706 observations × 79,800 functional connectivity edges, derived from the upper triangular portion of each 400 × 400 sliding-window FC matrix. In this formulation, each observation corresponded to a whole-brain FC pattern at a given time window, and each connection edge was treated as a variable. PCA was used to identify dominant patterns of FC reorganization over time and subjects. HR-FC edges were defined based on this PCA-based analysis ¹⁶. Specifically, all FC edges were ranked by the absolute value of their PC1 coefficients, and the top 25% were classified as HR-FC edges.

3.5.3. Permutation null test for between-network variance

To assess whether the observed network-specific variance in ALFF modulation was greater than expected by random variation, a permutation-based null test was performed. In each permutation, dALFF PCA coefficients were randomly shuffled across ROIs while preserving the original number of ROIs within each intrinsic network. Network-wise summary statistics and the resulting between-network variance were recomputed for each permuted dataset using the same procedures applied to the original data.

This process was repeated 10,000 times to generate a null distribution of between-network variance. The observed variance was compared against this null distribution, and statistical significance was defined as the proportion of permutations yielding variance values equal to or greater than the observed value. This approach tests whether the network-specific organization of dALFF PCA coefficients reflects meaningful structured network-level effects rather than random variation across ROIs.

4. Results

4.1. Acute severe hypoxia elicits temporally dissociated responses across modalities

We compared global and temporal responses across multiple modalities during a 3-min acute severe hypoxia challenge (Fig. 1). Among the measured physiological variables, end-tidal O₂ (P_ETO₂) exhibited the earliest and most pronounced change at hypoxia onset and during recovery (Fig. 1b), decreasing from a baseline of 109.9 ± 4.7 mmHg to 35.3 ± 1.7 mmHg immediately prior to reoxygenation. In contrast, SpO₂ showed a slower temporal response (Fig. 1c), declining from 98.3 ± 1.3% to 69.3 ± 8.7%. Behavioral performance, indexed by the number of commission errors, exhibited an even more delayed response following hypoxia onset (Fig. 1d).

Figure 1.

Phase-structured multimodal responses to an acute severe hypoxia challenge (F_iO₂ = 7.7%).

(a) Experimental timing of hypoxic stimulation.

(b) End-tidal oxygen partial pressure (P_ETO₂).

(c) Peripheral oxygen saturation (SpO₂).

(d) Commission errors as behavioral deterioration.

(e) Bulk BOLD signal.

(f) Functional connectivity (FC) strength.

(g) Amplitude of low-frequency fluctuations (ALFF; 0.01–0.1Hz).

(h) Amplitude of high-frequency fluctuations (AHFF; 0.1–0.25Hz). Panels (d), (f), (g), and (h) were estimated using a causal sliding-window approach (90 s window length).

(i) Z-score normalized and lightly temporally smoothed summary of P_ETO₂, FC, ALFF, and commission errors, illustrating their distinct temporal evolution. Baseline values (t < 180 s) were aligned to zero. Vertical dashed lines indicate hypoxia onset and offset.

(j) Conceptual segmentation of the experiment into baseline, hypoxia, and recovery periods, with hypoxia and recovery further divided into adaptation, decompensation, rebound, and stabilization phases based on dominant multimodal features observed across measures.

(k) Phase-wise dominant features summarizing characteristic changes in FC, ALFF, and behavioral performance.

Shaded regions represent ± standard deviation in panels (a-f) and ± standard error in panels (g-i) across subjects.

From the concurrently acquired fMRI BOLD time courses, we calculated four functional metrics: bulk BOLD signal, functional connectivity (FC) strength, ALFF, and AHFF (Fig. 1e–h). Despite being derived from the same underlying BOLD signal, these metrics exhibited distinct temporal dynamics during severe hypoxia. The bulk BOLD signal showed exponential-like changes following hypoxia onset and re-oxygenation. FC strength increased progressively throughout hypoxia and decreased after hypoxia offset. AHFF increased approximately linearly before reaching a plateau and subsequently declined during recovery. In contrast, ALFF exhibited a non-monotonic temporal pattern, distinguishing it from the other BOLD-derived measures.

Figure 1i summarizes the temporally aligned and z-score-normalized trajectories of P_ETO₂, FC strength, ALFF, and behavioral error. This comparison highlights both the temporal ordering and relative magnitude of changes across modalities. Based on distinct response characteristics relative to hypoxia onset and offset, the time course was operationally grouped into four phases: early hypoxia, late hypoxia, early recovery, and late recovery. Figure 1j and 1k schematically summarize this phase segmentation by marking the temporal boundaries on the task timeline and illustrating the dominant response characteristics within each period. Operationally, these phases correspond to adaptation (initial FC emergence without measurable increase in behavioral error), decompensation (progressive reduction in ALFF accompanied by increasing behavioral error during sustained severe hypoxia), rebound (transient overshoot in ALFF following re-oxygenation), and stabilization (return toward baseline dynamics).

4.2. Dynamic ALFF modulation is global yet exhibits network-specific structure

To assess the spatial structure of dynamic ALFF (dALFF) induced by acute severe hypoxia, we applied PCA to normalized dALFF time series across all 400 ROIs. As shown in Fig. 2a, the first principal component (PC1) accounted for 52.90% of the total variance, whereas PC2 accounted for 6.17%. Together, these two components explained 59.08% of the total variance, whereas each remaining component accounted for less than 3.40%, highlighting the dominant contribution of PC1 and a secondary, but meaningful contribution of PC2.

Figure 2.

Dynamic amplitude of low-frequency fluctuation (dALFF) exhibits a global, nonlinear, and network-specific response across acute severe hypoxia.

(a) Explained and cumulative variances of the first five principal components (PCs) derived from z-score–normalized ROI-level ALFF time series.

(b) Temporal profile of the first principal component (PC1), capturing the dominant phase-structured ALFF modulation across the experiment.

(c) Distribution of ROI-level PC1 coefficients across the 17 brain networks of the Schaefer atlas, illustrating network-specific contributions to the dominant ALFF mode.

(d) Permutation-based null distribution of between-network variance for PC1, demonstrating that the observed network structure exceeds chance level.

(e) Temporal profile of the second principal component (PC2), representing a secondary mode of ALFF modulation associated with decompensation and rebound phases.

(f) ROI-level PC2 coefficients across brain networks.

(g) Corresponding permutation-based null distribution for PC2.

In panels (b) and (e), shaded regions represent ± standard deviation across subjects, vertical dashed lines indicate hypoxia onset and offset, and red horizontal bars denote the hypoxia period.

The temporal profile of PC1 (Fig. 2b) closely resembled the global dALFF trajectory shown in Fig. 1g. PC1 coefficients were summarized across the 17 brain networks defined by the Schaefer atlas (Fig. 2c). All PC1 coefficients were positive, indicating a spatially coherent and globally synchronized ALFF modulation rather than localized or opposing effects. Nevertheless, the magnitude of these coefficients varied across networks. Permutation-based null testing revealed that the observed between-network variance of PC1 coefficients significantly exceeded chance levels (p < 0.0001; Fig. 2d), confirming the presence of structured network-specific heterogeneity within this global pattern.

Although PC2 explained a smaller proportion of variance (6.17%), it captured temporally specific fluctuations that were most prominent during the decompensation and rebound phases (Fig. 2e). In contrast to PC1, PC2 exhibited more differential network engagement, with both positive and negative coefficients across networks (Fig. 2f). Among networks, Default A showed the lowest median coefficient, whereas SomMotB showed the highest median coefficient, indicating opposing modulation in network involvement during these phases. Permutation testing again confirmed that this pattern also reflected meaningful network-specific structure rather than random spatial variation (p < 0.0001; Fig. 2g).

4.3. Functional connectivity reorganization aligns with pronounced ALFF modulation in a network-specific manner

To examine the relationship between changes in functional connectivity (Fig. 1f) and local brain activity (Fig. 1g) during acute severe hypoxia, we performed network-level analyses linking HR-FC edges to percent ALFF change (ΔALFF). Following our previous PCA-based framework ¹⁶, HR-FC edges were defined from dFC. For each ROI, we calculated the proportion of HR-FC edges (out of 400 possible connections) and the corresponding percent change in ALFF relative to baseline. These measures were then summarized at the network level using the Schaefer 17-network parcellation.

Figures 3a–d display the relationship between HR-FC ratio and ΔALFF across networks and across all four response phases. HR-FC ratios differed across networks, with the HR-FC set fixed across phases. Excluding Limbic A and B, DefaultA and SomMotB showed the highest HR-FC ratios (median [Q1, Q3]: 0.415 [0.283, 0.513] and 0.415 [0.305, 0.474], respectively), whereas the VisCent network exhibited the lowest HR-FC ratio (0.082 [0.000, 0.149]).

Figure 3.

Phase-dependent, network-specific coupling between functional connectivity and dynamic local activity.

(a-d) Network-wise relationship between changes in dynamic ALFF (ΔALFF, %) and the proportion of hypoxia-responsive functional connections (HR-FC ratio) across four response phases: adaptation (t = 220–240 s), decompensation (t = 340–360 s), rebound (t = 400–420 s), and stabilization (t = 550–570 s) to baseline (t = 160–180s). Each marker represents a canonical brain network based on the Schaefer 17-network parcellation, with error bars indicating median [Q1, Q3] across ROIs. Horizontal lines denote zero change in ΔALFF.

(e) Subject-level changes in functional connectivity strength (ΔFC) across phases for selected networks (DefaultA, SomMotB, and VisCent).

(f) Corresponding subject-level changes in dynamic ALFF (ΔALFF) across phases for the same networks.

Paired Wilcoxon signed-rank tests were used for network-wise comparisons (*p < 0.05; †0.05 < p < 0.1).

Taken together, these results demonstrate an association between network-level connectivity reorganization and local activity modulation, while indicating that even networks exhibiting pronounced connectivity reorganization can show divergent directions of local activity regulation. For example, during the decompensation phase, local activity markedly decreased in the Default A network, whereas a pronounced increase was observed in the SomMotB network.

Despite similar HR-FC ratios, DefaultA and SomMotB displayed markedly divergent ALFF responses, particularly during the decompensation phase (Fig. 3b). DefaultA showed the largest ALFF reduction among all networks (−18.4% [−28.8%, −9.9%]), whereas SomMotB exhibited a clear increase in ALFF (12.0% [−0.3%, 20.1%]). These opposing responses were contrasted with the adaptation phase, during which both networks showed modest ALFF increases (DefaultA: 6.3% [0.3%, 10.2%]; SomMotB: 7.2% [5.0%, 13.0%]). Thus, the decompensation phase was characterized by phase-specific divergence in ALFF modulation among networks.

The disparity persisted during the rebound phase following re-oxygenation (Fig. 3c). DefaultA showed a relatively modest ALFF increase (27.1% [11.5%, 45.6%]), whereas SomMotB demonstrated the largest ALFF increase (89.6% [75.4%, 103.1%]) among all networks. In contrast, the VisCent network exhibited comparatively modest ALFF changes during both the decompensation and rebound phases (−9.9% [−15.6%, −6.7%] and 26.4% [18.0%, 32.0%], respectively), consistent with its lower HR-FC ratio and indicating weaker ALFF modulation relative to DefaultA and SomMotB. During stabilization (Fig. 3d), ALFF increases were observed across most networks, but inter-network differences were attenuated compared with the rebound phase.

Subject-level analyses of the three networks (DefaultA, SomMotB, and VisCent) showed similar patterns across phases. Network-averaged FC strength increased during hypoxia (adaptation and decompensation phases), with sustained elevations into the rebound phase in DefaultA and SomMotB (Fig. 3e). ALFF changes reproduced the phase-specific divergence (Fig. 3f): DefaultA showed significant ALFF reductions during decompensation followed by moderate rebound increases, whereas SomMotB exhibited sustained ALFF levels during decompensation and significant increases during rebound. VisCent exhibited comparatively smaller ALFF changes. FC reorganization and ALFF modulation exhibited network-specific patterns during severe hypoxia, and increases in FC were not consistently accompanied by increases in ALFF.

4.4. Mild hypoxia induces limited temporal deviation in dynamic ALFF despite functional connectivity modulation

Figure 4 illustrates the global and temporal responses across multiple modalities during mild hypoxia (FiO₂ = 11.8%), which was administered continuously throughout the 10-min fMRI acquisition (Fig. 4a).

Figure 4.

Multimodal and network-level responses under acute mild hypoxia (F_iO₂ = 11.8%).

(a) Experimental timing of the mild hypoxia challenge.

(b) End-tidal oxygen partial pressure (P_ETO₂). P_ETO₂ decreased and stabilized at approximately 50 mmHg. The vertical dashed line (240 s) marks the time point corresponding to P_ETO₂ ≈ 53 mmHg, at which hypoxia-responsive functional connectivity (HR-FC) emerges.

(c) Peripheral oxygen saturation (SpO₂).

(d) Functional connectivity (FC) strength, showing a gradual increase beginning near the P_ETO₂ threshold.

(e) Amplitude of low-frequency fluctuations (ALFF; 0.01–0.1 Hz).

(f) Amplitude of high-frequency fluctuations (AHFF; 0.1–0.25 Hz). Panels (d)–(f) were estimated using a causal sliding-window approach (past 90 s).

(g) Subject-wise ΔFC across three representative networks (DefaultA, SomMotB, and VisCent), illustrating changes relative to the reference time point (t = 160–180 s) across time points.

(h) Subject-wise ΔALFF across the same networks. Overall changes were modest relative to the reference time point (t = 160–180 s); however, DefaultA showed smaller deviations from baseline compared with SomMotB and VisCent at multiple time points (†, paired Wilcoxon signed-rank test). Collectively, mild hypoxia induced reorganization of functional connectivity near the P_ETO₂ threshold while local fluctuation amplitudes remained relatively stable, particularly within the DefaultA network.

End-tidal oxygen (P_ETO₂) declined rapidly and exponentially following hypoxia onset (Fig. 4b), reaching 52.8 ± 3.0 mmHg at t = 240 s (vertical line) and gradually decreasing to 49.9 ± 3.6 mmHg by the end of the scan. Peripheral oxygen saturation (SpO₂) decreased more slowly (Fig. 4c), stabilizing at 81.1 ± 2.6%. These values remained substantially higher than those observed during severe hypoxia, consistent with a milder degree of oxygen reduction.

Hypoxia-responsive functional connectivity emerged around the P_ETO₂ critical point at t = 240 s (Fig. 4d) ¹⁶. However, unlike severe hypoxia, no clear adaptive and decompensatory phases were observed in dALFF (Fig. 4e). Relative to the monotonic increase observed under normoxia (Supplementary Fig. 1c), dALFF showed minimal temporal deviation from the reference interval. dAHFF showed no marked modulation, although a modest upward tendency was observed after t = 300 s (Fig. 4f).

Network-wise relative changes in FC and ALFF across subjects are shown in Figures 4g and 4h (reference interval: t = 160–180 s). Because the mild hypoxia protocol did not include discrete gas transitions, time points corresponding to those used in the severe condition (240, 360, 420, and 570 s) were adopted for direct comparison.

FC exhibited delayed enhancement following the P_ETO₂ critical point, with more apparent increases emerging after t = 360 s. However, the overall magnitude of FC change was smaller than that observed during severe hypoxia (cf. Fig. 3e), and network-specific differentiation was less pronounced (Fig. 4g).

In contrast, dynamic ALFF showed only limited temporal deviation relative to the reference interval. No consistent large-scale increase or decrease was observed across networks. Among the representative networks, DefaultA demonstrated relatively constrained deviation after t = 360 s compared with SomMotB and VisCent, remaining closer to the baseline level (Fig. 4h). Overall, these findings indicate that, despite progressive SpO₂ decline and measurable FC modulation, dynamic ALFF, particularly within DefaultA, remained comparatively stable during steady-state mild hypoxia.

5. Discussion

In this study, we investigated hypoxia-induced dynamic changes in spontaneous BOLD activity using ALFF with FC and interpreted these responses within the framework of canonical large-scale brain networks. Our findings revealed that, although both FC and ALFF highlighted largely overlapping dominant networks such as DefaultA and SomMotB, their temporal profiles and directional changes diverged. This dissociation suggests that ALFF reflects a complementary layer of functional dynamics, integrating both network-level reorganization and systemic oxygen fluctuations (reflected by SpO₂). Importantly, ALFF reduction exhibited greater temporal similarity to cognitive deterioration during hypoxia, indexed by changes in commission error rate, than did FC reorganization.

To integrate these observations across physiological, image-derived, and behavioral domains, we propose an oxygen budgeting (OB) framework that conceptualizes acute severe hypoxia as a system-level challenge of metabolic expenditure regulation (Fig. 5). In this framework, reductions in arterial oxygen partial pressure are accompanied by large-scale reorganization of FC, reflecting a coordinated network response to metabolic stress. Regional spontaneous neurovascular activity, indexed by ALFF, adapts within this reorganized network architecture, giving rise to the observed non-monotonic temporal modulation. Rather than simply mirroring oxygen availability, ALFF dynamics appear to reflect how limited metabolic resources are distributed across networks under constrained conditions. Cognitive deterioration may arise when sustained hypoxic stress overwhelms compensatory network and metabolic adaptations. From this perspective, hypoxic brain responses are not isolated regional failures but reflect coordinated and distributed network-level adaptations under constrained oxygen supply.

Figure 5.

Schematic illustration of the proposed oxygen budgeting framework for brain responses to acute severe hypoxia.

The schematic depicts a multi-layer conceptual model of brain responses under systemic metabolic constraint. Reductions in arterial oxygen partial pressure (PaO₂; approximated by P_ETO₂) provide an early physiological signal of reduced oxygen availability. Upon crossing a critical threshold, (1) large-scale functional connectivity reorganizes at the network level, reflecting a proactive pathway of network reconfiguration ¹⁶.

In parallel, progressive reductions in peripheral oxygen saturation (SpO₂) constrain cerebral oxygen delivery and are associated with (2) network-wise heterogeneous adjustments in cerebral oxygen metabolism ⁸.

(3) Local spontaneous neuronal activity, characterized in the present study, is positioned as an intermediate layer linking network reorganization with oxygen-dependent metabolic processes. Cognitive changes are conceptualized as downstream outcomes when sustained metabolic stress exceeds compensatory capacity.

Arrows indicate hypothesized directional relationships inferred from physiological coupling and temporal sequencing and do not imply strict causality. Here, “oxygen budgeting” denotes coordinated redistribution of multi-layer functional dynamics under metabolic constraint, providing a system-level interpretation of brain responses to acute severe hypoxia.

Consistent with this interpretation, ROI-wise ALFF changes observed in the present study paralleled cerebral metabolic rate of oxygen (CMRO₂) changes previously reported during the late phase of severe hypoxia in the same subjects ⁸. After matching ROI-level results across the two analyses and aggregating them at the network level, similar modulation patterns emerged, linking network-level ALFF dynamics with CMRO₂ changes (R² = 0.37; Supplementary Fig. 2). A related form of temporal dissociation was also observed during steady-state mild hypoxia, characterized by delayed enhancement in FC ¹⁶ alongside relative preservation of ALFF (Fig. 4). This pattern further suggests that ALFF modulation cannot be explained solely by systemic oxygen fluctuations but instead reflects coordinated network-level functional adjustments. In contrast, under normoxia, no notable network-level changes were observed in either FC or ALFF (Supplementary Fig. 3).

Prior PET–BOLD correspondence studies have shown that resting-state ALFF corresponds more with cerebral blood volume than with oxidative metabolic measures ²⁷. In the present study, ALFF modulation during severe hypoxic and recovery periods was not spatially uniform across the brain but varied systematically across networks and phases, arguing against a purely passive effect driven by global cerebral blood flow changes ⁸. By contrast, AHFF, reflecting higher-frequency BOLD fluctuations (Fig. 1h), exhibited relatively limited phase-dependent modulation, consistent with vascular pulsation and physiological fluctuations rather than neurovascular coupling. Together, these findings suggest that ALFF dynamics during acute hypoxia reflect network-specific neurovascular responses that are selectively organized rather than globally flow-driven. These effects should be interpreted as reflecting coordinated neurovascular adjustments under metabolic constraint.

The default mode network (DMN) is better understood not as a task-related functional unit, but as a high-energy-demand infrastructure supporting the brain’s intrinsic baseline state ²⁸. Given that a substantial portion of cerebral energy is dedicated to sustaining this spontaneous neural activity, the DefaultA-pronounced ALFF and FC changes observed during hypoxic decompensation phase likely reflect perturbation of the intrinsic baseline state rather than task-specific modulation. The concurrent decrease in ALFF alongside enhanced functional connectivity of DefaultA may represent complementary processes, reflecting downscaling of local spontaneous neural activity. Reduced local activity may limit the exposure of energetically costly synaptic processes to hypoxia-induced stress ²⁹. This process could potentially support system-level stability during oxygen limitation, albeit with measurable costs to behavioral performance.

Network-level analysis revealed relative increases in ALFF and FC within the SomMotB network during the hypoxic decompensation phase, a pattern distinct from other networks. In the 17-network parcellation ³⁰, SomMotB comprises ventral somatomotor regions including tongue-related areas, auditory cortex, and the secondary somatosensory cortex (S2), in contrast to the dorsal subnetwork (SomMotA), which is primarily associated with externally oriented motor-sensory processing. The right-lateralized dominance centered on the S2 region was observed, consistent with the proposed role of this region in interoceptive and viscerosensory integration ³¹. In the experimental context, hypoxia does not present as a salient sensory threat but instead manifests as an internal metabolic constraint. One possible interpretation is that networks involved in monitoring of internal physiological states remain relatively engaged under hypoxic stress. Whether the relative preservation in SomMotB network activity reflects sustained interoceptive or viscerosensory processing under acute severe hypoxic stress warrants further investigation.

Several limitations should be acknowledged. First, because both ALFF and FC were derived from BOLD signals, the observed changes may reflect a combination of neural, vascular, and systemic physiological influences, particularly under hypoxic conditions. Although P_ETCO₂ decreased during hypoxia, arguing against a simple hypercapnia-related vasodilatory explanation, the relative contributions of neural and non-neural factors cannot be fully disentangled in the present study. Second, the sample size was modest, and the findings therefore require replication in larger cohorts. Nevertheless, the within-subject graded hypoxia design and PCA-based identification of dominant patterns supported the detection of coherent group-level responses. Third, the network-level analysis based on the Schaefer parcellation enabled characterization of large-scale organizational trends, but may have obscured substantial heterogeneity within individual networks, where some regions potentially responding differently from the network-average pattern. Accordingly, the reported network-level effects should be interpreted as aggregate tendencies rather than spatially uniform responses within each network.

In summary, this study examined brain responses to acute hypoxia by characterizing alterations in local spontaneous activity, indexed by ALFF, alongside network-level changes in functional connectivity. Our findings suggest that network-based ALFF represents a distinct but complementary layer of functional dynamics temporally dissociated from functional connectivity, capturing local amplitude modulation that coexists with large-scale network reorganization. Although these measures capture distinct dimensions of BOLD dynamics, their hypoxia-related alterations were systematically linked, indicating that metabolic limitation concurrently reshapes both local spontaneous activity and large-scale network organization. Framed within the concept of oxygen as a metabolic budget, these multi-layered changes indicate that when oxygen availability is constrained, the brain does not respond uniformly but reorganizes its functional architecture across interacting levels. We refer to this integrated response as oxygen budgeting, reflecting coordinated redistribution of functional dynamics under metabolic limitation. Together, these findings provide important insights into the reorganization of brain networks under metabolic constraint.

This integrated response, which we refer to as oxygen budgeting, may fundamentally reflect a strategic shift between two competing demands: functional performance and structural stability. Given the brain’s exceptionally high baseline metabolic load ²⁸, largely dedicated to maintaining ionic homeostasis and synaptic transmission, sustained metabolic stress poses a direct threat to cellular integrity. In this context, the observed reductions in spontaneous activity and the reorganization of functional connectivity are not merely signs of regional failure or short-term behavioral optimization. Instead, they may represent an adaptive limitation of function aimed at preserving synaptic and network integrity under constrained energy supply. From this perspective, the hypoxic brain response reflects a form of neural resilience, where short-term functional compromise is traded for the protection of long-term structural viability, ensuring that metabolic resources are prioritized for survival over performance.