Calibration‐Free Estimation of Cerebrovascular Reactivity From Resting‐State BOLD fMRI Using Reconstructed Respiratory Fluctuations

ABSTRACT

Purpose

Cerebrovascular reactivity (CVR) provides an important index of vascular health and is conventionally quantified using a hypercapnic gas or breath‐hold challenge in conjunction with blood‐oxygen‐level‐dependent functional magnetic resonance imaging (BOLD‐fMRI). Such approaches require a dedicated extra scan and, for hypercapnia, specialized equipment for gas administration and external physiological recordings, limiting their applicability in large‐scale neuroimaging studies and clinical populations. To address this limitation, we introduce a calibration‐free and breath‐hold‐free framework for CVR estimation from resting‐state BOLD‐fMRI.

Methods

The method leverages a previously validated machine learning‐based respiratory variation (RV) reconstruction approach to recover respiratory dynamics directly from BOLD‐fMRI time series. The reconstructed RV is subsequently convolved with a respiratory response function and incorporated as a regressor in a voxel‐wise general linear model (GLM), yielding regression coefficients that serve as CVR estimates.

Results

Validation was performed on a cohort of 83 healthy young adults with ground‐truth CVR maps obtained from hypercapnic gas challenges. The proposed framework demonstrated spatial correspondence with measured CVR (mean correlation = 0.61), with the strongest performance observed in participants exhibiting greater respiratory variability (mean r = 0.72).

Conclusion

Collectively, these results establish a reliable, non‐invasive, and calibration‐free strategy for CVR mapping, enabling broader deployment in both research and clinical environments where gas‐challenge protocols and physiological monitoring are impractical.

1. Introduction

Cerebrovascular reactivity (CVR) reflects the ability of cerebral blood vessels to dilate in response to vasoactive stimuli, providing a crucial index of vascular health and autoregulation. Impaired CVR has been associated with a wide range of neurological disorders, including stroke, small vessel disease, cognitive decline, and Moyamoya disease [1, 2, 3], underscoring its importance as both a biomarker and a tool for understanding cerebrovascular pathophysiology.

Traditionally, CVR is measured using hypercapnic gas challenges in combination with blood‐oxygen‐level‐dependent functional magnetic resonance imaging (BOLD‐fMRI) [4]. Alternatively, breath‐hold paradigms can also be used to induce transient hypercapnia by increasing arterial carbon dioxide (CO₂) through voluntary apnea [5, 6]. Although breath‐hold approaches eliminate the need for specialized gas delivery systems, they depend heavily on subject compliance and can introduce variability in the resulting CO₂ response, particularly in pediatric or elderly populations. While both methods yield physiologically interpretable CVR maps, gas‐inhalation challenges remain the gold standard because they allow precise and reproducible control of the vasoactive stimulus. However, this requirement for specialized equipment, calibration, and monitoring limits their feasibility in large‐scale or vulnerable populations where controlled gas administration is impractical.

Resting‐state fMRI (rs‐fMRI) has emerged as an attractive alternative for CVR mapping, leveraging spontaneous fluctuations in arterial CO₂ as a natural contrast agent. When paired with end‐tidal CO₂ (P_ETCO₂) recordings, voxel‐wise regression of P_ETCO₂ against rs‐fMRI time series produces reliable CVR estimates [7, 8]. However, concurrent P_ETCO₂ measurements present challenges, as they necessitate specialized gas‐monitoring equipment and precise calibration, thereby increasing both the cost and logistical complexity of implementation. Furthermore, the use of nasal cannulas can be uncomfortable for participants and is susceptible to inaccuracies if subjects mouth‐breathe, move during scanning, or fail to adhere to breathing instructions [9].

To circumvent the reliance on P_ETCO₂ recordings, surrogate regressors such as global BOLD signals or respiration volume per time (RVT) have been proposed [10]. These approaches eliminate the need for specialized gas‐monitoring equipment while still enabling the derivation of CVR maps. Nonetheless, they remain constrained by significant limitations: global BOLD‐based estimates are highly susceptible to contamination from widespread neural activity and non‐vascular fluctuations, whereas RVT‐based methods depend on concurrent respiratory belt recordings that are frequently unavailable, improperly acquired, or corrupted [11, 12, 13]. More advanced strategies, including Global Regressor Signal Selection (GRSS) [14, 15], and GRSS‐informed deep learning frameworks [16], offer improved flexibility but ultimately still rely on surrogate regressors that may introduce systematic biases across populations, acquisition protocols, and scanner platforms.

These limitations underscore the need for a calibration‐free framework that derives CVR directly from resting‐state BOLD‐fMRI signals, without reliance on external physiological recordings, surrogate regressors, or restrictive modeling assumptions. In our previous work [12], we demonstrated that respiratory variation (RV) can be accurately reconstructed from resting‐state BOLD‐fMRI using a machine learning approach, enabling recovery of respiratory dynamics without auxiliary sensors. Building on this foundation, the present study introduces a fully physiological‐recording‐free framework for CVR estimation that leverages a machine learning model to reconstruct RV directly from resting‐state BOLD‐fMRI data and incorporates it into a general linear model (GLM) to estimate cerebrovascular responsiveness. We hypothesize that this calibration‐free framework will yield reliable CVR estimates that exhibit strong spatial correspondence with gold‐standard gas‐challenge measurements (e.g., correlation coefficients exceeding 0.6), thereby enabling broader application in both research and clinical settings.

2. Methods

2.1. Participants

This study consisted of data from 83 healthy young adults (mean age 39 ± 14 years; 45 females, 38 males) who gave informed consent to participate in this study. The inclusion criteria were no history of neurological disease, respiratory disease, no presence of MRI contraindications, and ability to complete a 90‐min MRI session that included wearing a face mask during a dual‐calibrated respiratory gas challenge protocol inside the scanner. Exclusion criteria included a history of diagnosed neurological diseases, and severe respiratory disorders requiring a home supply of oxygen. This study was approved by the Conjoint Health Research Ethics Board at the University of Calgary, written consent was obtained for each participant, and all procedures were in accordance with the Tri‐Council Policy Statement: Ethical Conduct for Research Involving Human Participants.

2.2. Data Acquisition

Data were acquired on a 3 T GE MRI (Discovery MR750) with a 32‐channel Nova Medical head coil. Each dataset included a high‐resolution 3D T₁‐weighted BRAVO anatomical image, a 10‐min rs‐fMRI timeseries sensitized to BOLD contrast, and a 20‐min multiband multi‐echo pseudo‐continuous arterial spin labelling (pCASL) scan [17]. The multi‐echo pCASL sequence was used in the present work to calculate the ground‐truth CVR maps derived from the hypercapnic segments of the dual‐calibrated gas challenge protocol, specifically using the BOLD‐weighted second echo of the acquisition rather than the perfusion‐weighted ASL component. The sequence parameters for the 3D T₁‐weighted BRAVO were: TR = 6.68 ms, TE = 2.94 ms, inversion time = 650 ms, voxels = 1 mm³, sagittal slices. For the BOLD rs‐fMRI sequence: TR = 1400 ms, TE = 29 ms, flip angle = 90°, in‐plane resolution = 3.5 × 3.5 mm, 64 × 64 matrix, voxels = 3.5 × 3.5 × 4 mm with 1‐mm slice gap, total duration = 609 s (435 volumes). These four‐dimensional datasets captured temporal fluctuations in blood oxygenation linked to both neural and vascular activity. The pCASL sequence parameters were: TR = 3028 ms, TE₁ = 10.5 ms, TE₂ = 29.8 ms, flip angle = 85°, multiband factor = 3, labelling duration = 1450 ms, PLD = 1000 ms, voxels = 3.5 × 3.5 × 4 mm with 1‐mm slice gap, no background suppression, total duration = 1211 s (400 volumes). To correct for distortions in the pCASL timeseries, a B₀ field map was acquired using a multi‐echo fast spoiled gradient recalled echo sequence with TR = 500 ms, TE_1–4 = 2.3/4.5/6.6/8.7 ms, flip angle = 30°, with spatial resolution matched to the pCASL sequence.

2.3. Hypercapnic Protocol

A dual‐calibrated gas protocol was implemented during the pCASL acquisition. This 20‐min protocol required the use of an MRI‐compatible breathing circuit to achieve mild hypercapnia (5% CO₂, 2 × 2‐min) and hyperoxia (50% O₂, 2 × 3‐min), interleaved with medical air [18, 19]. Inspired gases were delivered at 20 L/min. The fixed‐concentration automatic gas delivery system consisted of a Digital Flo‐Box, Mass Flow Controllers (Sierra Instruments, Monterey, CA), and Gemini O₂ and CO₂ Monitor and Gas Analyzer (CWE Incorporated). Breathing traces were continuously collected from a sampling port separate from the gas input to obtain end‐tidal partial‐pressure (P_ET) measurements of O₂ and CO₂. For the present analyses, CVR maps were calculated from the hypercapnia manipulations and P_ETCO₂ only.

2.4. Hypercapnia‐Derived CVR Maps

Only the volumes from the second echo sensitized to BOLD contrast from the multiband multi‐echo pCASL timeseries were used to calculate the ground‐truth CVR maps. First, all images were converted from DICOM to NIfTI. Distortion correction was then applied to the timeseries using the acquired B₀ field map and the Fieldmap toolbox [20, 21] in SPM12 (http://fil.ion.ucl.ac.uk/spm) running on MATLAB version R2021b. Rigid‐body registration for motion correction was performed using routines from the ASLtbx [22]. Linear and non‐linear registration between the BOLD volumes and the subject's T₁ image was performed using Advanced Normalization Tools [23] followed by spatial smoothing (5 mm FWHM). All fMRI analyses were performed on a per‐subject basis with all images in native space.

For first‐level modeling, the BOLD signal change to hypercapnia was modeled using FSL FEAT where a gamma‐variate function with 30‐s mean lag and standard deviation was implemented. This relatively broad gamma‐variate model was selected to capture the slow vascular response dynamics associated with sustained hypercapnic stimulation, including gradual CO₂ wash‐in and wash‐out effects, rather than the faster transient response typically modeled using conventional neural hemodynamic response functions. Motion parameters were included as regressors of no interest and a high‐pass filter (420 s) was implemented. Because a pCASL sequence was used, a regressor modeling tag‐control alternation (−0.5 and 0.5 every TR) was also implemented. The parameter estimate maps were used to obtain the fractional signal changes to hypercapnia. These maps were divided by the change in P_ETCO₂ from the hypercapnic challenge to obtain the final CVR maps. The change in P_ETCO₂ was calculated from the difference between baseline P_ETCO₂ (here this was the mean of 10 regular breaths prior to the onset of the first gas epoch) and the P_ETCO₂ at the end (the final 10 regular breaths) of the 5% CO₂ epoch. Because there were two CO₂ epochs, the mean of these two epochs was used. A block‐design GLM approach was selected for the hypercapnic reference maps because the gas challenge paradigm provided temporally well‐controlled CO₂ manipulations with sustained vasoactive epochs. In contrast, the resting‐state framework relied on spontaneous respiratory fluctuations with less predictable temporal structure, motivating the use of voxel‐wise lag optimization for the reconstructed RV regressor.

2.5. BOLD rs‐fMRI Data Preprocessing

The BOLD rs‐fMRI images were corrected for geometric distortions caused by B₀ inhomogeneity using the TOPUP tool in FSL [24]; corrected for head motion by the MCFLIRT tool in FSL [25]; and registered to 2 mm MNI space using Advanced Normalization Tools (ANTs) [26]. Following spatial normalization, the rs‐fMRI volumes were smoothed with a 5 mm full‐width‐at‐half‐maximum (FWHM) Gaussian kernel to improve signal‐to‐noise ratio (SNR). In addition, the hypercapnia‐derived CVR maps were also registered and normalized to the same MNI space to allow direct comparisons with the predicted CVR maps.

The RV estimation model used to reconstruct physiological regressors from the rs‐fMRI data was previously trained on the Human Connectome Project‐Young Adult (HCP‐YA) dataset, which includes high‐quality resting‐state fMRI data acquired with a repetition time (TR) of 0.72 s, along with physiological recordings from 1040 healthy adults aged 22–35 years. In that dataset, respiratory signals were continuously recorded using a pneumatic belt at 400 Hz, enabling the development of a robust machine learning framework capable of reconstructing RV waveforms directly from fMRI signals. Importantly, the framework was also validated on independent datasets acquired with different repetition times (TR = 1.5 s and TR = 3.0 s), demonstrating generalizability beyond the original HCP‐YA acquisition protocol [12].

2.6. Proposed CVR Estimation Method

Cerebrovascular reactivity was estimated using a GLM in which the reconstructed RV served as the physiological regressor. The RV waveform was reconstructed directly from rs‐fMRI data using our previously validated machine learning framework [12]. RV was defined as the standard deviation of the respiratory waveform within a 6‐s sliding window centered at each time point, consistent with prior physiological fMRI studies [27]. This framework employs one‐dimensional convolutional neural networks (1D‐CNNs) trained to capture temporal dependencies in BOLD fluctuations across multiple regions of interest (ROIs). The model inputs consisted of average BOLD signals from 90 functional ROIs [28], together with six estimated head motion parameters derived from retrospective image realignment. These inputs were segmented into overlapping windows of 65 TRs, enabling the network to leverage both past and future contextual information for each prediction. Each window produced a single RV estimate, and the outputs from all windows were concatenated to generate a continuous RV time series spanning the full duration of the scan.

The reconstructed RV time series was subsequently convolved with a respiration response function (RRF) [29] to model the hemodynamic effects of respiratory fluctuations on the BOLD signal. This RRF, derived empirically from single deep‐breath experiments, captures a biphasic response characterized by an early overshoot followed by a pronounced undershoot peaking at approximately 16 s. Mathematically, it is expressed as the difference of two gamma‐variate functions, reflecting the delayed and prolonged hemodynamic effects of CO₂‐mediated vasodilation. The resulting regressor ($\mathit{RV}*\mathit{RRF}$) was systematically shifted over a range of −10 to +40 s in 1.4‐s increments to account for voxel‐wise variability in vascular response timing. For each voxel, the lag that provided the best model fit was selected, and the optimally shifted regressor was entered into a voxel‐wise GLM. The voxel‐wise GLM model produced a CVR amplitude map, represented by the $\beta$‐coefficient, which quantifies the percent BOLD signal change per unit of reconstructed respiratory variation (%BOLD/unit RV). This map reflects the spatial distribution and magnitude of cerebrovascular responsiveness across the brain.

Consistent with prior work [7, 10], negative CVR values, considered physiologically implausible in healthy gray matter, were set to zero. This step reduces the influence of noise‐driven artifacts in white matter and ventricles, ensuring that reported CVR maps reflect vascular reactivity in plausible physiological ranges. The overall workflow of the proposed calibration‐free CVR estimation framework, including RV reconstruction, convolution, lag optimization, and voxel‐wise GLM analysis, is summarized in Figure 1.

FIGURE 1.

Calibration‐free CVR estimation framework using reconstructed RV. Subject's rs‐fMRI data were used to reconstruct the RV waveform directly from BOLD signals via a previously validated machine learning model [12]. The reconstructed RV was then convolved with a respiration response function to generate a regressor modeling respiration‐induced hemodynamic effects. This regressor was systematically time‐shifted to perform voxel‐wise lag optimization, yielding maps of response delay (Lag Time). The optimally shifted regressor was subsequently used in a voxel‐wise GLM to estimate the amplitude of CVR (Estimated CVR, expressed as %BOLD per unit RV). The resulting amplitude map represents the spatial characteristics of CVR in a calibration‐free manner.

2.7. Evaluation

The performance of the proposed CVR estimation framework was evaluated by comparing predicted CVR maps against hypercapnia‐derived reference maps. Evaluation was performed at the regional level using the 90 functional ROIs [28]. Within each ROI, CVR values were averaged, and correlations were computed between predicted and reference regional CVR values. For the representative subjects, additional ROI‐level linear regression analyses were performed between the reference and estimated CVR values. Pearson correlation coefficients (r) and regression slopes ($\beta$) were calculated to assess spatial correspondence and relative scaling between the two CVR measures. This ROI‐based approach provides a robust comparison of spatial patterns while reducing sensitivity to voxel‐wise noise.

For visualization purposes only, CVR amplitude maps (both predicted and reference) were scaled by their 98th‐percentile voxel value within the gray matter mask, in line with the methodology of Zvolanek et al. [10]. This percentile‐based scaling ensured consistent color limits across subjects and reduced the influence of noise‐driven outliers, while all quantitative analyses were performed on the unscaled ROI values.

All preprocessing and subsequent analyses were performed on the Advanced Research Computing (ARC) cluster at the University of Calgary, which provided two processors and 200 GB of RAM, ensuring efficient data handling.

3. Results

3.1. CVR Estimation Performance

Figure 2 presents representative examples from three subjects demonstrating the range of model performance, one with strong spatial correspondence (“good”), one moderate (“medium”), and one weaker agreement between estimated and reference CVR maps. Across subjects, the estimated CVR maps closely resembled those derived from hypercapnic gas challenges, exhibiting consistent spatial patterns of vascular reactivity in cortical and subcortical gray matter regions. The strongest correspondence was observed in large vascular territories such as the middle and posterior cerebral arteries, where both maps revealed similar gradients of reactivity. Minor discrepancies were noted primarily in deep gray matter and periventricular regions, where signal‐to‐noise ratios are typically lower.

FIGURE 2.

Comparison of gas challenge‐derived CVR maps (reference) and estimated CVR maps (RV‐GLM) for three representative subjects. Subjects 1, 2, and 3 correspond to cases with good, medium, and weaker spatial correspondence, respectively (Subject 1: r = 0.82; Subject 2: r = 0.64; Subject 3: r = 0.41). Difference maps were calculated as Estimated CVR—Reference CVR. For visualization, reference and estimated CVR maps were independently scaled by their 98th‐percentile voxel value within the gray matter mask. The difference maps are shown using percentile‐scaled units. Note that Subject 3 uses a wider difference‐map scale range due to larger deviations between estimated and reference CVR maps. Color bars represent relative CVR amplitudes expressed in arbitrary units proportional to %BOLD/unit RV.

To further examine the relationship between estimated and reference CVR values, ROI‐level regression analyses were performed for the three representative subjects shown in Figure 2 (Figure 3). Each point represents the mean CVR value within one of the 90 functional ROIs. Consistent with the voxel‐wise observations, strong, moderate, and weaker spatial correspondence were observed for Subjects 1, 2, and 3, respectively (Subject 1: r = 0.82, β = 2.28; Subject 2: r = 0.64, β = 1.31; Subject 3: r = 0.41, β = 1.01). The scatter plots did not reveal obvious evidence of distinct ROI populations. Instead, estimated and reference CVR values exhibited a continuous approximately linear relationship across ROIs, indicating that the proposed framework preserves the regional distribution of cerebrovascular reactivity despite differences in CVR scaling.

FIGURE 3.

ROI‐level comparison between reference and estimated CVR values for the three representative subjects shown in Figure 2. Each point represents the mean CVR value within one of the 90 functional ROIs. The red line indicates the least‐squares linear regression fit. Pearson correlation coefficients (r) and regression slopes ($\beta$) are reported for each subject. The scatter plots demonstrate positive spatial correspondence between estimated and reference CVR values across ROIs, while differences in $\beta$ reflect scaling differences between the relative CVR estimates derived from reconstructed respiratory variation and the gas challenge‐derived reference CVR maps.

Quantitative group‐level results (Figure 4) further confirmed the accuracy and reliability of the proposed calibration‐free framework. Across 83 participants, the correlations between predicted and gas challenge‐derived regional CVR values were consistently positive, with a mean correlation of 0.61 and a median correlation of 0.64. The distribution of correlation coefficients demonstrates generally positive spatial correspondence across the cohort, although substantial inter‐subject variability was observed. Higher agreement was primarily observed in participants exhibiting greater respiratory variability, whereas subjects with lower RV variability showed more moderate correspondence between estimated and measured CVR maps.

FIGURE 4.

Distribution of scan‐wise correlations between estimated and reference CVR maps across all participants. Each point represents one subject's correlation coefficient, with the boxplot summarizing the overall distribution. The mean correlation (0.61) is shown by the horizontal thick black line, and the median correlation (0.64) is highlighted by the white circle. The consistently positive correlations demonstrate strong spatial correspondence between estimated and reference CVR maps across individuals.

3.2. BOLD Temporal Characteristics and CVR Estimation Fidelity

To investigate the sources of intersubject variability in CVR estimation accuracy, Figure 5 presents representative grayplots of resting‐state BOLD signals from two participants, one exhibiting high estimation fidelity (r = 0.82) and another showing lower estimation fidelity (r = 0.41). Each grayplot depicts voxel‐wise BOLD signal fluctuations over time, arranged by spatial location, with the reconstructed RV waveform displayed above.

FIGURE 5.

Temporal characteristics of resting‐state fMRI signals underlying CVR estimation fidelity. Representative grayplots of resting‐state BOLD‐fMRI signals from two participants with (a) higher (r = 0.82) and (b) lower (r = 0.41) correspondence between estimated and gas challenge‐derived CVR maps. Each grayplot displays voxel‐wise signal fluctuations over time, with the reconstructed RV waveform shown above. Vertical black‐white banding patterns reflect respiration‐driven global BOLD modulations. In panel (a), orange arrows indicate pronounced vertical bands corresponding to large respiratory pattern changes, while green arrows highlight smaller but distinct respiratory fluctuations. In contrast, panel (b) shows minimal vertical banding, suggesting reduced respiration‐related modulation and a weaker physiological contribution to the BOLD signal.

In the participant with higher estimation fidelity (Figure 5a), the grayplot demonstrates pronounced vertical black‐white banding patterns, reflecting temporally coherent global BOLD signal modulations. These oscillatory structures correspond closely to changes in the reconstructed RV waveform and are characteristic of respiration‐induced fluctuations that propagate across widespread brain regions. The presence of such well‐defined temporal regularities indicates a strong respiratory contribution to the BOLD signal, thereby facilitating more accurate RV reconstruction and consequently, more reliable CVR estimation. This observation is consistent with our previous findings [12], which showed that RV reconstruction accuracy improves when distinct respiratory pattern changes are present.

In contrast, the participant with lower estimation fidelity (Figure 5b, r = 0.41) exhibits limited or absent vertical banding, suggesting reduced temporal coherence and a weaker respiratory influence on the global BOLD signal. The diminished respiration‐related modulation results in a less informative physiological signature, reducing the model's ability to accurately reconstruct RV dynamics and leading to lower spatial correspondence between estimated and reference CVR maps.

To further test whether the variability of respiratory dynamics contributes to differences in CVR estimation accuracy, all participants were analyzed based on the standard deviation of their reconstructed RV waveform, which quantifies the magnitude of breathing pattern fluctuations. A higher RV standard deviation reflects greater variability in breathing depth and rate, whereas a lower value indicates more stable respiration.

Across all 83 participants, the median RV standard deviation was used as a threshold to classify subjects into two groups:

• High RV variability (n = 41): participants exhibiting greater respiratory fluctuations.

• Low RV variability (n = 42): participants exhibiting more stable breathing patterns.

As shown in Figure 6a, a strong positive correlation was observed between RV variability and CVR estimation accuracy (r = 0.78), indicating that greater variability in respiratory dynamics enhances the model's ability to recover vascular responses. Consistently, participants in the high‐variability group demonstrated significantly higher spatial correspondence between estimated and reference CVR maps (mean r = 0.72) than those in the low‐variability group (mean r = 0.49), as illustrated in Figure 6b. An independent‐samples t‐test revealed a significant difference between the two groups, t(81) = 10.15, p < 0.001. These findings confirm that increased respiratory variability provides a richer endogenous signal for calibration‐free CVR estimation. Collectively, these results suggest that the proposed framework provides more reliable CVR estimates in subjects exhibiting sufficient spontaneous respiratory variability, while estimation confidence may be reduced in participants with weaker endogenous respiratory fluctuations.

FIGURE 6.

Relationship between respiratory variability and CVR estimation accuracy. (a) Across all participants, the standard deviation of the reconstructed RV waveform (RV standard deviation) showed a strong positive correlation with CVR estimation accuracy (r = 0.78), indicating that greater respiratory variability enhances vascular signal detectability. (b) Distribution of correlation coefficients (r) between estimated and gas challenge‐derived CVR maps for high and low RV variability groups. Participants were divided by the median RV standard deviation, with the high‐variability group (n = 41) showing significantly higher correspondence with reference CVR maps (mean r = 0.72) than the low‐variability group (n = 42; mean r = 0.49; p < 0.001, independent‐samples t‐test).

In addition, we quantified the frequency of vertical banding patterns observed in the grayplots, representing temporally coherent fluctuations in the global BOLD signal. On average, participants in the high‐variability group exhibited 2.8 pronounced vertical bands (orange in Figure 5), compared with only 0.7 in the low‐variability group. These visual findings reinforce the quantitative results, indicating that greater respiratory variability enhances the physiological information encoded in the resting‐state BOLD signal. Smaller, yet distinct vertical bands (green in Figure 5) were consistently present across nearly all scans, suggesting that low‐amplitude respiration‐related modulations are a ubiquitous feature of the resting‐state BOLD time series. The presence of both large and small banding structures across subjects highlights the wide range of respiration‐induced temporal regularities in resting‐state fMRI data and supports the observation that greater variability in respiratory dynamics corresponds to improved accuracy of calibration‐free CVR estimation.

4. Discussion

4.1. Physiological Basis and Interpretation of Findings

The positive correspondence between estimated and measured CVR maps demonstrates that spontaneous BOLD fluctuations inherently encode physiologically meaningful information about vascular responsiveness. Respiration is a primary driver of low‐frequency BOLD variability, acting mainly through oscillations in arterial CO₂ that modulate cerebral blood flow and oxygenation [29]. By reconstructing the RV waveform directly from resting‐state BOLD signals, the proposed framework effectively isolates this physiological component, enabling voxel‐wise estimation of CVR without the need for external recordings. The consistently positive correlations observed across participants confirm that the learned mapping between regional BOLD fluctuations and underlying respiratory dynamics generalizes across individuals and scanning sessions.

Temporal analyses of the grayplots (Figure 5) provide mechanistic insight into intersubject differences in estimation fidelity. Scans exhibiting pronounced respiration‐related banding patterns, reflecting temporally coherent global BOLD modulations, produced higher reconstruction accuracy and consequently, more reliable CVR estimates. This finding supports the notion that the detectability of vascular oscillations in resting‐state fMRI depends on the magnitude and temporal regularity of spontaneous respiratory fluctuations. When these fluctuations are weak or inconsistent, the physiological imprint within the BOLD signal diminishes, reducing model sensitivity. Accordingly, the present framework may provide more confident and physiologically informative CVR estimates only in subjects exhibiting sufficient spontaneous or induced respiratory variability.

To further assess whether reduced reconstructed RV variability reflects true physiological variability or attenuation introduced by the reconstruction model, supplementary analyses were performed using the HCP‐YA dataset containing concurrent respiratory recordings (Supporting Information, Section S1). These analyses demonstrated a strong positive association between reconstructed and measured respiratory variability, supporting the interpretation that low RV variability primarily reflects underlying physiological characteristics of the resting‐state recordings, although the reconstruction model may introduce a modest reduction in dynamic range.

These results align with prior work showing that respiration‐ and CO₂‐driven components of the global BOLD signal are key determinants of robust CVR estimation during rest [7, 10]. In the absence of sufficient respiratory variability, the endogenous vasoactive signal becomes limited, reducing the ability to reliably estimate CVR. This relationship is evident in the positive association between RV variability and CVR estimation accuracy (Figure 6a) and further supported by the group‐level analysis of RV variability (Figure 5b), where participants with greater fluctuations in breathing depth and rate exhibited significantly higher correspondence between estimated and reference CVR maps (p < 0.001). Collectively, these findings demonstrate that physiological variability, often regarded as a confounding factor in fMRI, can serve as an informative endogenous probe of cerebrovascular function when appropriately modeled.

Notably, pronounced fluctuations in respiratory patterns are frequently observed even among healthy young adults scanned at rest, despite explicit instructions to maintain consistent breathing rate and depth. Prior analyses of the HCP‐Young Adult dataset by Power et al. [13], and Lynch et al. [30], demonstrated that transient deep breaths and serial burst‐like breathing events commonly occur, inducing global BOLD signal transients and coherent vascular oscillations. Similar variability patterns were also observed in the respiratory recordings analyzed in the present supplementary analyses (Supporting Information, Figure S2). These findings underscore the inherent physiological variability present in so‐called “resting‐state” conditions, reinforcing the suitability of our calibration‐free framework for such acquisitions. Importantly, respiratory irregularities and altered respiratory control have been reported more frequently in aging populations [31], and neurodegenerative disorders, including Parkinsonian syndromes, dementia with Lewy bodies, and Alzheimer's disease [32, 33, 34]. Prior studies have documented respiratory dysrhythmia, sleep‐disordered breathing, impaired ventilatory control, and reduced respiratory muscle function in these populations, potentially reflecting brainstem and autonomic dysfunction associated with neurodegeneration. In these populations, spontaneous fluctuations in breathing depth and rate may serve as an enhanced endogenous vasoactive stimulus, potentially enabling even more robust and informative CVR estimation. However, pathological alterations in neurovascular coupling or cerebrovascular regulation may also modify the relationship between respiration and the BOLD response, warranting further validation of the proposed framework in aging and clinical cohorts.

Building on this, participants could be instructed to perform brief, controlled breathing alterations, such as a few paced deep breaths during a portion of the resting‐state fMRI scan, to increase CO₂ fluctuations and improve the reliability of calibration‐free CVR estimation. In scans exhibiting low spontaneous RV variability, such brief breathing modulations may help enhance the endogenous vasoactive signal and improve the robustness of CVR estimation. This concept is consistent with the intermittent breath‐modulation approach introduced by Liu et al. [35], where short paced‐breathing blocks were embedded within an otherwise resting scan to enhance E_TCO₂ variability and improve CVR sensitivity. As shown in our data, higher respiratory variability was consistently associated with more accurate CVR estimation (Figure 6a), suggesting that reconstructed RV variability may serve as a useful internal confidence metric for calibration‐free CVR estimation. In the present cohort, participants with RV standard deviation values above approximately 0.15 generally exhibited improved estimation performance (r > 0.6); however, this value should be interpreted as an empirical observation specific to the current dataset rather than a universal threshold. The estimated variability of the reconstructed RV waveform may therefore provide a practical indicator of the physiological signal strength available for CVR prediction in each scan. Accordingly, RV variability may be useful as a quality‐control metric for identifying scans in which calibration‐free CVR estimates should be interpreted with greater caution. Together, these strategies may help standardize resting‐state CVR estimation protocols and extend their applicability to a wider range of populations and clinical settings.

4.2. Comparison With Existing Calibration‐Free Methods

Traditional calibration‐free CVR approaches have primarily relied on surrogate regressors such as the global BOLD signal, principal component‐derived components, or respiration volume per time (RVT) when available [10, 15, 16]. While these methods can approximate the spatial distribution of CVR, they exhibit several important limitations. The global BOLD signal conflates vascular and neuronal sources of variance, obscuring regional specificity and introducing systematic biases in areas of high neural activity. RVT‐based regressors, although physiologically grounded, require concurrent respiratory belt recordings that are often unavailable in resting‐state datasets [12]. Moreover, RVT is an imperfect index of respiration‐induced CO₂ fluctuations. As demonstrated by Addeh et al. [36], RVT can fail to detect deep or irregular breaths and exhibits unstable behavior during breath‐holding, leading to inaccurate modeling of BOLD signal variations in such conditions.

Compared with RVT‐based calibration‐free approaches [10], our framework achieved slightly lower correlation with gas challenge‐based reference maps. However, it is important to note that Zvolanek et al. [10], analyzed breath‐hold fMRI data, which provide a strong and explicit CO₂ modulation, whereas our method was evaluated on resting‐state acquisitions that contain subtler natural fluctuations, leading to inherently lower SNR. Furthermore, their study compared the RVT regressor with the end‐tidal P_ETCO₂ signal rather than with gas challenge‐derived CVR maps. Because both RVT and P_ETCO₂ reflect respiration‐dependent fluctuations, a high correlation between them is expected and reflects their shared physiological basis, rather than direct equivalence in quantifying cerebrovascular reactivity.

Golestani et al. [7], introduced one of the earliest quantitative resting‐state CVR frameworks, employing simultaneous recordings of P_ETCO₂, respiratory volume, and cardiac pulsation to model voxel‐wise BOLD responses via parametric deconvolution. Their approach provided a rigorous and physiologically grounded means of estimating cerebrovascular reactivity without explicit gas challenges. Nevertheless, the framework required comprehensive physiological monitoring and extensive preprocessing, including removal of both high‐ and low‐frequency respiratory and cardiac components prior to regression. While this denoising strategy was designed to isolate vascular responses, it may have inadvertently eliminated physiologically meaningful fluctuations that carry information about CO₂‐driven vascular dynamics. Because respiration‐ and cardiac‐related signals encode variations in arterial CO₂ and cerebral blood flow, their complete suppression can reduce sensitivity to endogenous vascular reactivity. In contrast, the present framework leverages these fluctuations as informative features rather than artifacts, transforming them into a natural vasoactive contrast for calibration‐free CVR mapping.

By reconstructing respiratory dynamics directly from BOLD signals, the proposed framework eliminates dependence on external recordings or surrogate global metrics. The machine‐learning model flexibly captures nonlinear mappings between distributed BOLD fluctuations and respiratory variation, enabling consistent performance across acquisition protocols and scanner settings. The resulting CVR maps are physiologically interpretable, data‐driven, and fully calibration‐free, facilitating application to a broad range of archival and prospective fMRI datasets that lack auxiliary physiological measurements.

4.3. Methodological Implications and Clinical Potential

From a methodological perspective, the proposed framework provides a bridge between physiological modeling and data‐driven inference. By combining machine learning with established physiological response functions, it integrates the interpretability of model‐based CVR mapping with the adaptability of data‐driven reconstruction. The resulting CVR maps maintain high spatial fidelity while reflecting realistic vascular dynamics. The inclusion of voxel‐wise lag optimization was particularly important because cerebrovascular response timing can vary substantially across brain regions and individuals, influencing the accuracy of CVR amplitude estimation.

Importantly, because this approach requires only standard resting‐state fMRI acquisitions, it may facilitate application to large public datasets such as the Human Connectome Project, UK Biobank, or ADNI. However, estimation performance is expected to depend on the degree of spontaneous respiratory variability present within each scan. Future work should investigate standardized quality‐control metrics and adaptive acquisition strategies to improve robustness in low‐variability datasets while preserving the practicality of calibration‐free CVR mapping.

Although the CVR values estimated by the proposed framework are expressed in relative rather than absolute physiological units, this limitation does not compromise their interpretability. In most applications, the spatial pattern of cerebrovascular responsiveness is of primary importance, and accurate relative contrasts are sufficient for identifying regional vascular deficits. Consistent with previous work [7, 10], each CVR map can be normalized by its global mean to enable cross‐subject comparison without loss of generalization. This normalization preserves regional variability while mitigating global scaling differences, thereby facilitating meaningful assessment of cerebrovascular heterogeneity across populations.

Clinically, calibration‐free CVR mapping holds promise for assessing vascular health in populations where gas challenges or physiological monitoring are impractical, such as pediatric, elderly, or critically ill patients. The ability to estimate relative CVR from resting‐state scans may aid early detection of cerebrovascular impairment, longitudinal monitoring of disease progression, and evaluation of therapeutic interventions. Furthermore, because RV reconstruction operates directly on BOLD data, the same framework could be adapted to task‐based fMRI or multimodal imaging (e.g., PET/MRI), enabling broader applications in neurovascular and neurodegenerative research.

ROI‐level regression analyses (Figure 3) provided additional insight into the relationship between estimated and reference CVR values. Although the regression slopes varied across representative subjects (β = 2.28, 1.31, and 1.01), the estimated and reference CVR values exhibited positive linear relationships in all cases. Furthermore, the scatter plots did not reveal obvious evidence of distinct ROI populations or systematic regional bias. Instead, the estimated and reference CVR values followed a continuous approximately linear relationship across ROIs, suggesting that the proposed framework preserves the spatial distribution of cerebrovascular reactivity despite differences in amplitude scaling. These scaling differences are expected because the proposed framework derives CVR from spontaneous respiratory fluctuations during resting‐state fMRI rather than from controlled hypercapnic gas challenges. Consequently, the estimated CVR values should be interpreted primarily as relative measures of vascular responsiveness.

4.4. Limitations

Despite the strengths of the proposed approach, several limitations should be acknowledged. First, the CVR values obtained through this framework are expressed in relative rather than absolute physiological units. While normalization facilitates cross‐subject comparison, it precludes direct quantification of cerebrovascular sensitivity in absolute terms (e.g., %BOLD/mmHg CO₂). Nevertheless, as emphasized by previous studies, accurate spatial distribution is often more critical than absolute magnitude for identifying regional CVR differences. Future work could address this limitation by calibrating reconstructed RV amplitudes against concurrently measured physiological signals in a subset of participants.

Second, the current model was trained and validated on healthy young adults. Although the results demonstrate strong generalizability within this demographic, physiological and hemodynamic characteristics may differ substantially in older individuals or clinical populations. Additionally, alterations in neurovascular coupling, respiratory regulation, or cerebrovascular physiology in aging and neurological populations may influence the relationship between respiratory fluctuations and BOLD responses, potentially affecting reconstruction accuracy and CVR estimation performance. Validation across diverse cohorts, including those with impaired autoregulation, altered respiratory control, or vascular pathology, will be essential for broader clinical adoption. Furthermore, because the present study was limited to healthy participants, the sensitivity and specificity of reconstructed RV‐based CVR maps for detecting regional cerebrovascular abnormalities remain to be established through comparison with hypercapnia‐derived CVR maps in clinical populations.

Third, differences in acquisition parameters between datasets (e.g., TR, scan length, and SNR) may influence reconstruction accuracy. Although the frequency‐domain consistency observed in Figure 5 suggests robustness to such factors, systematic evaluation across multi‐center datasets is warranted. Additionally, the model currently assumes a fixed respiration response function (RRF), which may not fully capture regional or subject‐specific variations in vascular delay. Incorporating adaptive or individualized hemodynamic modeling could further enhance sensitivity.

Fourth, the voxel‐wise lag estimates derived from the reconstructed RV regressor were not directly validated against ground‐truth delay measurements obtained from hypercapnic gas challenges. Although lag optimization was incorporated to improve temporal alignment between the physiological regressor and BOLD responses, cerebrovascular delay can vary substantially across individuals and may be altered in pathological conditions. Future work should therefore evaluate the physiological accuracy and interpretability of reconstructed RV‐based lag estimates relative to established delay‐mapping approaches.

Finally, despite the proposed breathing‐modulation strategies and confidence metrics, some scans may still exhibit insufficient respiratory variability for reliable calibration‐free CVR estimation. In such cases, the endogenous physiological signal may be too weak to support robust RV reconstruction and subsequent CVR mapping. Additional physiological recordings or conventional gas‐challenge approaches may therefore still be required in a subset of subjects.

4.5. Future Directions

Incorporating multimodal physiological priors, such as cardiac activity or refined motion‐derived parameters, could further improve reconstruction accuracy and enable simultaneous estimation of cardiac and respiratory influences on BOLD fluctuations. The integration of more advanced architectures, such as convolutional‐recurrent or transformer‐based networks, may enhance temporal context modeling, particularly for longer or lower‐SNR scans.

Another important direction for future research is the evaluation of the proposed framework across different brain parcellation schemes and spatial scales. Although the present study used a 90‐region functional atlas to facilitate ROI‐level comparisons, alternative atlases may capture different aspects of cerebrovascular organization and could influence regional CVR estimates. Future work should therefore investigate the robustness of the observed spatial correspondence across functional, anatomical, and multi‐scale parcellations, while also exploring strategies to improve agreement in CVR amplitude between RV‐based estimates and gas‐challenge‐derived reference maps. In addition, future studies incorporating a subset of participants with concurrent respiratory belt recordings could facilitate both calibration of reconstructed RV amplitudes and direct comparison between CVR maps derived from reconstructed and measured respiratory signals, providing further validation of the proposed framework. Furthermore, applying the method to patient populations will provide critical insight into its diagnostic utility. For example, calibration‐free CVR mapping could help identify early vascular dysfunction in small vessel disease, Alzheimer's disease, or traumatic brain injury, where traditional gas challenges are not feasible. Finally, combining reconstructed RV‐based CVR estimation with dynamic functional connectivity analysis may reveal how vascular responsiveness modulates brain network organization, advancing our understanding of neurovascular coupling in health and disease.

5. Conclusion

This study introduces a calibration‐free framework for estimating CVR directly from resting‐state BOLD‐fMRI, eliminating the need for an extra scan, external physiological recordings, or gas‐challenge calibration. By reconstructing RV directly from resting‐state BOLD‐fMRI signals and incorporating it into a physiological GLM, the framework enables voxel‐wise CVR mapping and robust ROI‐based evaluation. The reconstructed RV exhibited realistic respiratory frequency characteristics, and the resulting CVR maps demonstrated positive spatial correspondence with those obtained from hypercapnic gas challenges, with stronger performance observed in participants exhibiting greater respiratory variability.

These findings demonstrate that spontaneous BOLD fluctuations inherently contain sufficient physiological information to recover both respiratory dynamics and cerebrovascular responsiveness. The proposed approach provides a practical and physiologically interpretable alternative for assessing vascular health in large research cohorts and clinical populations where traditional calibration procedures are not feasible, particularly in scans exhibiting sufficient spontaneous respiratory variability. Future developments integrating multimodal physiological priors, individualized hemodynamic modeling, and validation across diverse populations will further enhance its robustness, generalizability, and clinical applicability.

Funding

A.A. was supported in part by an Eyes High Doctoral Scholarship and the Alberta Innovates Doctoral Scholarship. G.B.P. acknowledges support from the Canadian Institutes of Health Research (FDN‐143290) and the Natural Sciences and Engineering Research Council of Canada (RGPIN‐03880). M.E.M. acknowledges support from the Natural Sciences and Engineering Research Council of Canada through a Discovery Grant (RGPIN‐03552) and an Early Career Researcher Supplement (DGECR‐00124).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Relationship between measured and reconstructed respiratory variability across the HCP‐YA dataset. Each point represents one scan, with the x‐axis showing the standard deviation of the measured RV waveform and the y‐axis showing the standard deviation of the reconstructed RV waveform. The dashed black line represents the identity line (y = x), and the red line shows the linear fit. The fitted regression demonstrated a positive relationship between measured and reconstructed RV variability (slope = 0.86).

Figure S2: Representative respiratory signals from scans with low and high respiratory variability. The top three panels show examples of respiratory traces from participants with lower respiratory variability standard deviation (RV SD) values, reflecting relatively stable breathing patterns throughout the scan. The bottom three panels show examples with higher RV SD values, characterized by larger fluctuations in breathing depth and rate and more irregular respiratory dynamics. These examples illustrate the substantial intersubject variability in spontaneous respiration observed during resting‐state fMRI acquisitions.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.