Physiological Confounds in BOLD‐fMRI and Their Correction

ABSTRACT

Functional magnetic resonance imaging (fMRI) has opened new frontiers in neuroscience by instrumentally driving our understanding of brain function and development. Despite its substantial successes, fMRI studies persistently encounter obstacles stemming from inherent, unavoidable physiological confounds. The adverse effects of these confounds are especially noticeable with higher magnetic fields, which have been gaining momentum in fMRI experiments. This review focuses on the four major physiological confounds impacting fMRI studies: low‐frequency fluctuations in both breathing depth and rate, low‐frequency fluctuations in the heart rate, thoracic movements, and cardiac pulsatility. Over the past three decades, numerous correction techniques have emerged to address these challenges. Correction methods have effectively enhanced the detection of task‐activated voxels and minimized the occurrence of false positives and false negatives in functional connectivity studies. While confound correction methods have merit, they also have certain limitations. For instance, model‐based approaches require externally recorded physiological data that is often unavailable in fMRI studies. Methods reliant on independent component analysis, on the other hand, need prior knowledge about the number of components. Machine learning techniques, although showing potential, are still in the early stages of development and require additional validation. This article reviews the mechanics of physiological confound correction methods, scrutinizes their performance and limitations, and discusses their impact on fMRI studies.

This graphical abstract depicts physiological confounds in BOLD‐fMRI from respiratory and cardiac sources, categorized by frequency, and highlights their effects on brain regions through various mechanisms. The figure contrasts model‐based correction methods with data‐driven techniques, outlining the strengths and limitations of each approach in enhancing signal fidelity and reducing spurious connectivity.

Abbreviations

ABCD

Adolescent Brain Cognitive Development

aCompCor

Anatomical Component‐Based Confound Correction

BMI

body mass index

BOLD

blood oxygen level dependent

CBF

cerebral blood flow

CBV

cerebral blood volume

CMRO₂

cerebral metabolic rate of oxygen

CNN

convolutional neural network

CO₂

carbon dioxide

Comp

CorComponent‐Based Confound Correction

CSF

cerebrospinal fluid

CRF

cardiac response function

CRV

cardiac rate variation

dHb

deoxyhemoglobin

DORK

Dynamic Off‐Resonance in k‐Space

ECG

electrocardiograms

EPI

echo‐planar imaging

FIX

FMRIB's ICA‐based X‐noiseifier

fMRI

functional magnetic resonance imaging

fNIRS

functional near‐infrared spectroscopy

GRE

gradient‐recalled echo

GLM

general linear model

Hct

hematocrit

HRV

heart rate variation

HCP‐A

Human Connectome Project in Aging

HCP‐D

Human Connectome Project in Development

HCP‐Y

AHuman Connectome Project in Young Adults

HRF

hemodynamic response functions

ICA

independent component analysis

M

calibration constant

MCA

middle cerebral artery

ML

machine learning

nROI

confound regions‐of‐interest

NMR

nuclear magnetic resonance

OEF

oxygen extraction fraction

PCA

principal component analysis

P_ETCO₂

partial pressure of end‐tidal CO₂

RetroKCor

Retrospective k‐space Correction

RetroICor

RETROspective Image CORrection

RIPTiDe

Regressor Interpolation at Progressive Time Delays

ROI

regions of interest

RV

respiratory variation

RRF

Respiration Response Function

RVT

respiratory volume per time

rs‐fMRI

Resting‐State Functional Magnetic Resonance Imaging

sLFOs

systemic low‐frequency oscillations

SENSE

Sensitivity Encoding

SNR

signal‐to‐noise ratio

SPM

Statistical Parametric Mapping

tCompCor

Temporal Component‐Based Confound Correction

sICA

Spatial Independent Component Analysis

TE

time of echo

TR

time of repetition

tSTD

temporal standard deviation

tICA

Temporal Independent Component Analysis

TCD

transcranial doppler

VASO

Vascular‐Space‐Occupancy

VCG

vectorcardiograms

1. Introduction

Functional magnetic resonance imaging (fMRI) is a powerful neuroimaging modality that provides in vivo measurements of brain activity, via neurovascular coupling, with superior spatial resolution compared to other functional imaging techniques [1]. It has been extensively used in behavioral experiments to study brain function associated with specific tasks or conditions, and during resting conditions to examine the brain's intrinsic functional architecture [2, 3]. The majority of fMRI experiments are based on the blood oxygen level dependent (BOLD) contrast [4]. The BOLD effect occurs as a result of changes in the magnetic susceptibility of blood when the concentration of deoxyhemoglobin (dHb) in the brain's capillaries and venules varies, inducing field gradients within and around vessels and altering the MR signal intensity.

The concentration of dHb is influenced by cerebral blood flow (CBF), cerebral metabolic rate of oxygen (CMRO₂), and cerebral blood volume (CBV). These hemodynamic parameters, in conjunction with baseline physiological states including blood hematocrit (Hct), resting state oxygen extraction fraction (OEF), and CBV, intricately influence the BOLD signal's amplitude [5, 6, 7, 8, 9, 10]. The complex interplay between these elements can lead to similar BOLD responses despite varied changes in CBF, CMRO₂, and CBV for the same baseline physiological state. For example, Blockley et al. [1] showed that both visual stimulation and breath‐holding could provoke similar BOLD responses, though their underlying mechanisms diverge—visual stimulation affects both CBF and CMRO₂, whereas breath‐holding primarily impacts CBF. This highlights the importance of considering the specific hemodynamic and metabolic factors that contribute to the BOLD signal when interpreting both task‐evoked and resting‐state BOLD signal changes.

Subtle physiological variations, such as low‐frequency fluctuations in respiratory depth and rate, as well as heart rate, are known to influence cerebral physiological parameters and head motion and, consequently, the amplitude of the BOLD signal [11, 12, 13]. Significant correlations have been demonstrated between these physiological variations and the BOLD signal [14]. These physiological fluctuations fall within the frequency range of resting‐state BOLD oscillations [0.01 0.15] Hz [15] (approximately 0.03 Hz for respiration [11, 12] and 0.04 Hz for cardiac [13]), potentially leading to artificially inflated functional connectivity measures or the appearance of spurious connectivity patterns where true BOLD connectivity does not exist [16, 17]. Notably, these low‐frequency variations in respiration and heart rate serve as primary drivers of systemic low‐frequency oscillations (sLFOs). sLFOs represent a broader category of systemic hemodynamic signals, incorporating not only respiratory and cardiac influences but also blood pressure oscillations (Mayer waves) and vasomotor fluctuations, which collectively overlap with and obscure the true neuronal low‐frequency BOLD signal, necessitating careful disentanglement for accurate functional connectivity analysis [17].

Efforts to minimize head movements substantially reduce movement‐related signal changes; however, respiration‐induced changes and the intrinsic pulsatility of blood flow remain present in BOLD‐fMRI signals. Respiratory cycles, primarily around 0.3 Hz [18], lead to thoracic movements that instigate changes in the magnetic field within the head, causing respiratory‐dependent phase shifts in the MR image [19, 20]. Separately, cardiac cycles, at a higher frequency of about 1 Hz [21], induce cardiac pulsatility that results in small, yet significant movements in brain tissue. Furthermore, pulsations within large blood vessels due to cardiac‐induced pressure variations lead to minor displacements in the surrounding tissue [21]. Respiratory and cardiac motions occur at frequencies significantly higher than the low‐frequency (< 0.1 Hz) BOLD fluctuations usually investigated in resting‐state fMRI studies. The commonly used repetition time (TR) in BOLD echo‐planar imaging (EPI) (~0.5–3 s) results in the aliasing of high‐frequency respiratory and cardiac confounds into the low‐frequency domain [22]. Consequently, these fluctuations could be misinterpreted as neural activity‐related BOLD oscillations, underscoring the importance of distinguishing true neural signals from physiological confounds.

Physiological confounds exhibit non‐stationary behavior, with amplitudes that vary considerably over time, posing significant challenges to accurate modeling within fMRI data analysis. Such variations are not uniform across different demographics and may exhibit substantial intra‐ and inter‐individual variability. This issue is particularly pronounced in older and clinical populations, where disrupted neurovascular coupling further complicates the comparison of BOLD signals across cohorts [8]. The increasing volume of literature on the impact of physiological variations on fMRI data, as evidenced by publication trends on Scopus.com from 1994 through 2024 (see Figure 1), highlights the escalating scholarly engagement with these complexities.

FIGURE 1.

Trends in published articles on the impact of physiological confounds in fMRI research. This figure illustrates the number of publications sourced from Scopus.com, using the search terms “fMRI,” “fMRI physiological confounds,” and “fMRI physiological noise” within the title, abstract, and keywords of the articles. It presents a comparative analysis between the total fMRI studies and those considering physiological confounds. The data reveals a growing but still limited scholarly focus on these confounds, as evidenced by the relatively small proportion of publications considering them compared to the overall number of fMRI studies. This discrepancy highlights a critical gap in current research methodologies, indicating that the majority of fMRI research has not yet systematically considered physiological confounds.

The pivotal challenge of effectively addressing physiological confounds in fMRI datasets has catalyzed the need for innovative research methods, leading to the development of three primary techniques: filtering approaches [22], model‐based approaches that incorporate additional physiological recordings [11, 12, 13, 14, 18, 23, 24], and data‐driven methods [25, 26]. Addressing physiological confounds during fMRI data preprocessing is important for the reliability and validity of results. Yet, there remain several limitations, such as the selection of an appropriate confound correction method [27], quality of externally recorded data [28, 29, 30, 31, 32], associated costs, and the interpretation of results. We aim to provide a comprehensive review of the confounds, existing physiological confound correction methods in fMRI studies and discuss associated limitations, with the goal of improving the accuracy and reliability of fMRI data analyses. In this review paper, we classified and named physiological confounds as (1) low‐frequency respiratory confound that occurs ~0.03 Hz, (2) low‐frequency cardiac confound that occurs 0.04 Hz, (3) high‐frequency respiratory confound or respiratory‐related motion that occurs ~0.3 Hz, and (4) high‐frequency cardiac confound or cardiac‐related motion that occurs ~1 Hz (The low‐frequency cardiac confound [0.04 Hz] reflects slow fluctuations in heart rate, while the high‐frequency cardiac‐related motion [~1 Hz] corresponds to faster pulsatile motion driven by each heartbeat). It's important to note that these rates may vary with age and disease, potentially affecting their impact and detection in fMRI studies. Exploring the role of extraneous variables such as age and gender on the BOLD signal is not the focus of the present review; however, this has been explored in prior work [33].

2. Mechanism of BOLD Effect

The BOLD effect is the result of two primary factors: a biophysical factor, wherein oxygenated hemoglobin displays diamagnetic properties in contrast to the paramagnetic nature of dHb [34], which leads to the creation of magnetic field distortions within and around blood vessels, thereby diminishing the MR signal [6, 35]; and a physiological aspect, where brain activation is marked by a decline in both the local OEF and the corresponding levels of dHb [6, 35, 36, 37, 38].

Upon delivery to the tissues, arterial blood is normally almost completely saturated with oxygen. In a resting state, the OEF is approximately 40%, as blood circulates through the capillary beds [6]. This process results in a substantial production of dHb within the capillaries and venous vessels. The paramagnetic properties of dHb generate magnetic field gradients near red blood cells and the surrounding tissue, which consequently shortens the T₂ and $T_2^{*}$ relaxation times and lowers the MR signal compared to a state without dHb.

When performing a task, active regions of the brain exhibit increased CMRO₂, leading to a localized rise in production of dHb. To meet the need for additional oxygen, nearby arterioles dilate, causing an increase in local blood flow. However, increases in CBF are much greater than increases in CMRO₂, leading to a decrease in the OEF and a net reduction in the concentration of dHb. This disproportionate coupling of CBF and CMRO₂ constitutes the foundational mechanism underlying the neurovascular coupling induced BOLD response in the brain [39]. Further complexity is introduced by variations in post‐arterial CBV, which influence the total volume of blood containing dHb, and by inducing a volume exchange effect. The volume exchange effect occurs when an increase in intravascular blood volume during activation compresses or displaces the surrounding extravascular space, altering the distribution of water protons, impacting local magnetic susceptibility gradients, and contributing to the observed BOLD signal changes [1, 40]. In a healthy brain, the change of CBF outweighs the alterations in the CMRO₂ and CBV, ultimately resulting in a net decrease of dHb concentration during increased neuronal activity [10]. With less dHb, the susceptibility of the blood moves closer to the susceptibility of the surrounding tissue. The $T_2^{*}$ becomes longer, leading to an increase in the signal measured with a gradient‐recalled echo (GRE) pulse sequence by several percent [37].

The Davis model (Equation (1)) [7] furnishes a mathematical construct that captures the nuances of extravascular BOLD signal fluctuations, elucidating the relationship between cerebral hemodynamic changes and their resultant effects on MR signal amplitude:

$$\frac{∆S}{S_0}=M\left[1-{\left(\frac{{\mathit{\text{CMRO}}}_2}{{\mathit{\text{CMRO}}}_{2_0}}\right)}^{\beta }{\left(\frac{\mathit{CBV}}{{\mathit{CBV}}_0}\right)\left(\frac{\mathit{CBF}}{{\mathit{CBF}}_0}\right)}^{-\beta }\right]$$ (1)

where M, the calibration constant, is determined by the magnetic field strength through the parameter k, the pulse sequence through the time of echo (TE), and the baseline physiological state via baseline CBV, baseline OEF, and the Hb concentration in the blood. This relationship is expressed as follows:

$$M=k\times \mathit{TE}\times {\mathit{CBV}}_0\times {\left({\mathit{OEF}}_0\times \left[\mathit{Hb}\right]\right)}^{\beta }$$ (2)

The exponent $\beta$ varies between 1 and 2, reflecting the field strength dependent non‐linear dependence of T ₂* on the concentration of dHb within a tissue sample [41]. Figure 2 offers a comprehensive visual representation of how various scan and physiological parameters influence the BOLD‐fMRI signal. It highlights the signal variations due to different scan settings and stimulus‐induced modulation of CBF, CBV, CMRO₂, dHb, and OEF, as well as physiological fluctuations like respiration and heart rate. Additionally, Figure 2 demonstrates that both task‐induced changes and physiological confounds, specifically breath‐holding in this example, can produce similar BOLD signals through distinct mechanisms [1, 42, 43], with the resultant BOLD signal changes appearing similar despite significant differences in the underlying alterations in CBF and CMRO₂.

FIGURE 2.

Illustration of the BOLD‐fMRI signal influenced by various scan and physiological parameters. (a) Shows the dependence of the BOLD response on changes in CBF, modulated by the ratio of fractional CBF to CMRO₂ changes, denoted by n, and the baseline dHb scaling factor, represented by M (panel a, left). It also illustrates how initial magnetization (M₀) (panel a, middle) and $T_2^{*}$ relaxation times (panle a, right) further influence the BOLD signal. (b) Illustrates the general process by which neural activity, whether evoked by an external task (e.g., a stimulus) or arising spontaneously, as in resting‐state fMRI, leads to changes in the BOLD signal. The diagram includes contributions from neuronal activity and confounding factors such as respiratory and cardiac signals, ultimately influencing the BOLD response observed over time. (c) Compares the relative changes in CMRO₂, CBF, and BOLD responses between task‐based stimuli and a breath‐holding condition. It highlights that task‐induced BOLD signals involve both CMRO₂ and CBF, whereas the breath‐holding BOLD response is driven predominantly by changes in CBF alone, producing a similar BOLD signal through a different physiological pathway.

3. Origin of Physiological Confounds

This section investigates physiological confounds affecting the BOLD signal through cerebral hemodynamics and magnetic field strength, as described by the deoxyhemoglobin dilution model [41].

3.1. Low‐Frequency Respiratory Confound

Carbon dioxide (CO₂) is as a potent vasodilator, and its presence in arterial blood can dilate arterial blood vessels, leading to increased CBF and CBV [44]. As a by‐product of aerobic glucose metabolism in tissues, CO₂ plays a crucial role in neurovascular coupling due to its vasodilatory effect. However, changes in respiratory rate and depth can also directly affect arterial CO₂ levels, leading to CBF alterations and subsequent BOLD signal fluctuations not associated with neuronal activity. Hyperpnea, characterized by deep or rapid breathing, increases CO₂ expulsion and decreases the partial pressure of CO₂ (PCO₂) in arteries, which in turn reduces CBF and the BOLD signal. Conversely, hypopnea or apnea, indicated by shallow, slow, or paused breathing, restricts CO₂ release, leading to increased arterial PCO₂, raised CBF, and an increased BOLD signal [12, 43, 45].

The relative contribution of intravascular BOLD signal changes to the overall BOLD signal modulation is potentially substantial, as evidenced by the quadratic relationship between the $T_2^{*}$ value of blood and the concentration of dHb. Notably, at magnetic field strengths below 3.0 Tesla, the intravascular component emanating from the venous structures constitutes a significant proportion of the BOLD signal—estimated at approximately 57% at a field strength of 1.5 Tesla, and around 36% at 3.0 Tesla [5]. CBV changes are strongly associated with CBF changes, as described by Grubb's law: $\mathit{CBV}~{\mathit{CBF}}^{\alpha }$[44], where $\alpha$ is a constant, with a reported value of 0.2 for venous CBV changes [46]. Increased venous CBV decreases the BOLD signal both by raising dHb concentration, which attenuates the signal, and by displacing tissue water, thereby reducing the tissue's contribution to the signal, a mechanism leveraged in Vascular‐Space‐Occupancy (VASO) [47] imaging to detect CBV changes. Consequently, CO₂‐driven alterations in CBF will impact CBV. Due to this volume effect, low‐frequency variations in breathing rate and depth can modulate the BOLD signal, introducing a non‐neuronal component to its measurement.

Research has shown variable effects on CMRO₂ with alterations in arterial CO₂ levels, with some studies reporting increases while others indicate decreases [10, 48]. Chen and Pike [49] specifically investigated both hypercapnic and hypocapnic conditions in a study with 10 subjects, modulating end‐tidal CO₂ partial pressure (P_ETCO₂) from −6 to +9 mmHg relative to a normocapnic baseline of approximately 40 mmHg, and observed no significant changes in CMRO₂. Given that the arterial CO₂ variation during an fMRI scan is approximately 1.1 mmHg [12], which is within the studied $∆P_{\mathit{ET}}{\mathit{CO}}_2$ range of −6 to +9 mmHg, the authors concluded that low‐frequency variations in breathing rate and depth do not significantly impact the BOLD signal through changes in CMRO₂.

Wise et al. [12] reported associations between P_ETCO₂, a proxy for arterial CO₂ levels, and BOLD signal variability, with oscillations of $\pm 1.1$ mm Hg in P_ETCO₂ corresponding to changes of $\pm 0.12\%$ in the BOLD signal in gray matter. The P_ETCO₂ variations, typically around 0.03 Hz, match the frequencies used in many behavioral fMRI paradigms [11, 12, 23], suggesting a potential for unaccounted confounds in fMRI data analysis [24, 50, 51]. In their investigation, Wise et al., used Transcranial Doppler (TCD) to measure the blood velocity of the middle cerebral artery (MCA) as an indirect indicator of tissue CBF, based on the assumption of a constant arterial diameter. They linked BOLD signal fluctuations primarily to CO₂‐driven changes in CBF but also noted regional variations in the BOLD response to CO₂ [12]. Wise et al. [12] suggested that this variability could indicate underlying differences in metabolic processes, capillary network density, and vascular regulatory mechanisms. These initial findings point to the need for further detailed analysis to understand the effects of CO₂ fluctuations on BOLD signal dynamics and the roles of different physiological contributors. For more accurate CBF quantification, arterial spin labelling (ASL) is preferred over TCD due to its precision [52]. The study also emphasizes the importance of considering CBV changes, particularly the impact of volume exchange between blood vessels and surrounding tissue, in the context of neurovascular coupling.

3.2. Low‐Frequency Cardiac Confound

Low‐frequency oscillations in heart rate play a significant role in the BOLD‐fMRI signal [13, 14, 53]. Studies have shown that low‐frequency fluctuations in heart rate, inherent in systemic cardiovascular dynamics, are linked to oscillations in cerebral hemodynamics. Katura et al. [53] conducted a quantitative assessment utilizing transfer entropy to elucidate the interrelationship between low‐frequency oscillations in cerebral hemodynamics—specifically changes in hemoglobin concentration—and systemic cardiovascular parameters such as heart rate and arterial blood pressure. Their findings suggest that approximately 20% of the variability in low‐frequency oscillations in cerebral hemoglobin concentration can be ascribed to variations in heart rate. The oscillations of hemoglobin concentration, believed to be conveyed through the impact of heart rate on CBF and CBV, influence the BOLD signal due to alterations in dHb concentration.

Shmueli et al. [13] examined the temporal and spatial patterns of BOLD signal changes associated with heart rate variations, reporting significant correlations. These time shifts correspond to the delay between heart rate fluctuations and their effect on the BOLD signal, with negative correlations in gray matter at time shifts of 6–12 s, likely reflecting the physiological lag in the vascular response to heart rate changes, and positive correlations at shifts of 30–42 s, which may indicate a different phase of the hemodynamic response. Their analysis suggests that low‐frequency heart rate variations are a potential source of physiological confound in BOLD‐fMRI, affecting the signal through their modulation of CBF and CBV. Furthermore, including time‐shifted cardiac rate timecourses as regressors in their models demonstrated a reduction in physiological confound variance, potentially increasing the statistical power of fMRI studies. The findings of Shmueli et al. [13] imply that heart rate fluctuations, especially those at low frequencies, must be accounted for to accurately interpret BOLD‐fMRI data and separate neuronal activation‐related changes from physiological confounds.

Several studies suggest that the same mechanisms underlying the correlation between low‐frequency CO₂ fluctuations and the BOLD signal may also influence the association between heart rate and the BOLD signal [13]. Notably, both CO₂ changes and heart rate variations affect similar brain regions, highlighting a shared physiological basis [13, 14]. The baroreflex, a key autonomic mechanism that regulates blood pressure, plays a crucial role in this interplay by adjusting heart rate and vascular resistance in response to fluctuations in arterial pressure. Through this regulation, the baroreflex influences respiratory patterns and CO₂ levels, which in turn affect CBF and the BOLD signal. However, the relationship between heart rate changes and the BOLD signal is intricate, influenced by a web of factors including sleep [54], respiration [55], blood pressure, and nervous system activity [56]. Future research should aim to elucidate these relationships further, potentially leading to more refined fMRI analysis techniques that can effectively account for such physiological variability.

3.3. High‐Frequency Respiratory Confound

Respiratory motion‐induced distortion in fMRI images is a multifaceted phenomenon rooted in the intrinsic properties of lung tissue and the mechanical act of breathing. Respiratory motion induces susceptibility variations in the lungs due to their lower magnetic susceptibility relative to surrounding tissues, leading to alterations in the induced magnetic field (B₀) within the brain. Consequently, this motion affects the magnetic field's spatial distribution throughout the body and head, causing fluctuations in resonant frequency and periodic structures in the Fourier domain of the image, which manifest as distortions in the phase of the MRI signal [19, 57]. To a lesser extent, respiratory motion also contributes to ghosting and blurring artifacts, which result from inconsistencies in phase increments between odd and even k‐space lines, thereby degrading image quality and impacting fMRI signal fidelity [58]. Noll and Schneider's seminal study was the first to rigorously examine the effects of subjects' respiration on fMRI data [57]. In their experiments, they examined the response at different echo times and determined that the size of respiratory phase variations increases approximately linearly with echo time. They concluded that the respiratory phase variations are likely due to resonant frequency variations. For example, they reported that a 1 Hz shift in the resonant frequency (0.016 ppm at 1.5 T) caused a 14° shift in the phase of the received signal at TE = 40 ms [57]. The size of these changes in the brain varies in range 0.15 Hz at the superior edge of the brain to 1 Hz in the temporal lobes (2° to 14° at a 40 ms TE) [57]. Interestingly, these phase changes were not confined to the brain; similar alterations were detected in phantoms positioned above the shoulder, suggesting that respiration affects the magnetic field distribution rather than the MR properties of the brain tissue itself [57]. This conclusion was corroborated by Brosch et al. [19], through experiments with a mechanical model that simulates human respiratory mechanics using a manikin. The system features an inflatable mechanism that mimics lung movements, causing the solution inside the manikin to displace, thereby emulating the abdominal expansion and contraction observed in human breathing. This setup enables the study of how respiratory‐induced magnetic field fluctuations affect MRI signals, while isolating these effects from other physiological influences [19].

The study by Bollmann et al. [59] employs concurrent field monitoring to analyze and correct field fluctuations in fMRI data, effectively disentangling physiological and hardware‐related confounds. By utilizing an array of NMR (nuclear magnetic resonance) field probes—small, highly sensitive sensors that measure local magnetic field variations in real‐time—the researchers recorded uniform and gradient magnetic field components during fMRI acquisition. This approach enabled spectral separation of physiological field fluctuations from hardware‐induced perturbations. While mechanical models like the one developed by Brosch et al. provide valuable insights into respiratory‐induced magnetic field changes, real‐time field monitoring is essential to quantify and correct these fluctuations in vivo. However, physiological fluctuations are not the only source of field instabilities in fMRI. Scanner‐related effects, such as magnet drift and gradient imperfections, also contribute to signal fluctuations. To further isolate these effects, the researchers conducted experiments using a phantom, a controlled, non‐biological object that allows for the study of field fluctuations in the absence of physiological influences such as breathing or cardiac activity. The phantom scans provided a baseline measurement of hardware‐related field instabilities, revealing that the dominant confounds in fMRI were primarily caused by magnet drift and gradient imperfections. By comparing these results with in vivo scans, the study demonstrated that breathing‐related field fluctuations were relatively small in magnitude compared to scanner‐induced variations. Although breathing‐induced fluctuations contribute less to overall field instability compared to hardware‐related effects, they remain an inherent and uncontrollable physiological factor in fMRI. Unlike hardware instabilities, which can be mitigated through field stabilization techniques, respiratory variations are subject‐dependent and dynamic, making them more challenging to correct retrospectively.

Respiratory cycles are relatively high frequency (~0.3 Hz), compared to the low‐frequency (< 0.1 Hz) BOLD fluctuations under analysis in resting‐state fMRI [32]. However, due to the long TR of traditional resting BOLD EPI (~2 s), respiratory confounds at the primary frequency are aliased into this low‐frequency range [16]. This aliasing causes respiratory confounds to appear as low‐frequency fluctuations that may be mistaken for neural activity‐related BOLD oscillations. Additionally, the frequency of susceptibility fluctuations caused by abdominal volume changes during respiration can vary depending on a variety of factors. These include the participant's breathing rate, the depth of their breaths, specific imaging parameters used, and individual physiological characteristics such as weight, height, body mass index (BMI) [32, 60], and sex [32] which may alter the volume and movement of the lungs and, consequently, the magnetic field distribution.

Building on this, research by Duerst et al. [61], utilizing NMR field probes in high‐field structural $T_2^{*}$‐weighted imaging at 7 T, demonstrated that respiratory‐induced magnetic field perturbations are more pronounced in individuals with higher BMI and in inferior brain slices located closer to the lungs. Their findings highlight a direct relationship between respiratory‐induced susceptibility variations and the volume of tissue involved in respiration, emphasizing the influence of body morphology on the magnitude of these distortions. Notably, the implementation of real‐time field control effectively mitigated these perturbations through dynamic compensation of respiratory field fluctuations via higher‐order shimming. These results suggest that analogous susceptibility‐induced field shifts observed in structural imaging are likely to extend to fMRI acquisitions, where they contribute to phase encoding shifts and image distortions. These effects may be quantitatively assessed by estimating phase encoding shifts and susceptibility‐induced image distortions in EPI, where the extent of shifts can be computed based on the bandwidth per pixel.

Measuring the susceptibility of living tissues presents a significant challenge in model development due to the inherently small susceptibilities and the heterogeneous nature of imaging volume, especially pronounced at air‐tissue interfaces, such as the ear canal, sinuses and nasal cavity [62]. These distortions are further exacerbated by thoracic motion, particularly in regions like the inferior surface of the brain and temporal lobes [18, 19, 57]. Thoracic motion further impacts the BOLD signal by inducing disturbances in steady‐state free precession of long T₂ species (e.g., CSF) during single‐shot EPI sequences [63]. While models can be developed to estimate susceptibility such as quantitative susceptibility mapping (QSM) [64, 65], accounting for changes in respiration among other effects in individual scans remains impractical. This complexity, compounded by the dynamic nature of respiratory‐induced distortions, underscores the necessity for adaptive imaging parameters. These are acquisition and reconstruction settings—such as echo time, repetition time, slice orientation, or dynamic shimming—that can be dynamically or periodically adjusted in real time to accommodate changing physiological conditions and minimize artifacts. Future research should focus on these adaptive approaches to significantly enhance the accuracy of fMRI data interpretation, allowing for more precise differentiation between true neural activity and artifacts introduced by respiratory motion.

3.4. High‐Frequency Cardiac Confound

Most fMRI studies rely on the assumption that the brain is a static physical system that undergoes no structural changes during an experiment. However, evidence from previous studies on brain physiology during the cardiac cycle have suggested that this is not the case. Poncelet et al. [66] showed that the brain has complex motion in response to blood pulsation, with lateral and cephalocaudal velocities as high as 1.5 mm/s in the thalami. In addition, parenchymal excursions up to 0.5 mm were found to occur in temporal synchrony with systole [66]. Vessel pulsation, cerebrospinal fluid movement, and tissue deformation are all associated with the cardiac cycle, and produce fMRI signal variance [21, 67, 68].

Given the finite capacity of the cranial volume, the infusion of blood instigates a dynamic interplay among the space demands of the blood volume, cerebrospinal fluid (CSF) pools, and brain tissue. With the onset of the systolic phase in the cardiac cycle, a rapid surge in pressure within the cerebral vasculature triggers an intracranial pressure wave that propagates along the cerebral arterial tree in a fraction of a second [69, 70]. Consequently, the arteries dilate to accommodate the incoming blood within the confines of the cranial cavity, initially in the frontal lobe and subsequently in the more posterior regions of the brain. Some studies have noted blood volume changes at the capillary level, with expansion of the arterioles and compression of the venous capillaries during systole [71]. The force exerted on the parenchyma from the expanding vasculature leads to inward expansion of the brain during systole and forces CSF down through the foramen magnum [71]. There is also uniform bulk motion of the large brain regions during systole and during diastole slowly returning to their initial shape [66]. Studies examining this motion have reported that the central structures such as the diencephalon and brain stem move in the caudal direction while the peripheral regions such as the major cerebral lobes and posterior cerebellar hemisphere move in the cephalic direction [72].

Hemodynamically driven pulsation of the brain can cause highly pulsatile phase shifts, which in turn result in motion artifacts whose intensity is larger than the activation signals [67, 68]. In a preliminary report by Glover and Lee [67], highly pulsatile phase shifts that coincide with the cardiac cycle were observed. These variations can manifest as distortions or shifts in the image because fMRI images are sensitive to differences in phase as mentioned. Previous studies showed reduced sensitivity due to cardiac‐induced variation in the BOLD signal is greater in specific areas, typically near major blood vessels and CSF pools [18]. Because these counfound‐induced signal changes can be equal to or greater than those produced by neuronal activity, the ability to detect signal changes due to neural activation would be decreased in these specific regions.

3.5. Summary of Physiological Confounds

Physiological confounds in BOLD‐fMRI, summarized in Figure 3, originate from respiratory and cardiac fluctuations, influencing the BOLD signal through distinct mechanisms. Low‐frequency respiratory confounds (~0.03 Hz) modulate arterial CO₂ levels, altering CBF and CBV, thereby causing global BOLD signal fluctuations, particularly in gray matter and regions sensitive to blood flow changes. High‐frequency respiratory confounds (~0.3 Hz) stem from respiratory motion‐induced magnetic field (B₀) variations, impacting areas near temporal lobes and air‐tissue interfaces. Low‐frequency cardiac confounds (~0.04 Hz) reflect heart rate oscillations that influence hemoglobin concentration, primarily affecting gray matter and cerebral arteries. High‐frequency cardiac confounds (~1 Hz), driven by blood pulsation and tissue displacement during the cardiac cycle, generate localized BOLD signal changes near large blood vessels, CSF‐filled spaces, and brain parenchyma. These mechanisms and their regional impacts are further explored in Section 9, which provides a detailed investigation of the brain's regional sensitivity to these physiological fluctuations.

FIGURE 3.

Physiological confounds in BOLD‐fMRI. The figure categorizes physiological confounds into respiratory and cardiac domains, outlining their characteristic frequencies, underlying mechanisms, and regional impacts on the BOLD signal. It highlights the necessity for targeted correction strategies to improve the accuracy of fMRI data analysis.

4. Measurement of Physiological Parameters in fMRI

In fMRI research, model‐based methods for confound correction critically depend on concurrent physiological recordings to effectively identify and mitigate specific artifacts while preserving the neural signal of interest. Physiological recordings commonly encompass measurements of respiratory movements and photoplethysmogram signals, which are routinely integrated into MR units. Such data are essential for correcting both low‐frequency and high‐frequency physiological fluctuations. Specifically, the measurement of P_ETCO₂ through capnography is a valuable additional metric for quantifying and correcting the effects of low‐frequency respiratory variations on the BOLD signal. This section delves into the complexities involved in obtaining accurate physiological measurements, highlighting the technical and methodological challenges inherent in this process.

4.1. Respiration Rate and Depth

Respiration rate and depth are critical physiological parameters monitored during fMRI to mitigate respiratory artifacts. These measurements are typically obtained using respiration belts—non‐invasive devices that enable continuous monitoring throughout the scanning procedure. A respiration belt, equipped with a stretchable band encircling the subject's chest or abdomen, incorporates a transducer that converts respiratory movements into a digital signal. The signal is subsequently analyzed to adjust for respiratory‐related artifacts in the fMRI data. The efficacy of respiration belts is contingent upon several factors including the correct positioning of the belt, calibration of the pressure sensor, and the consistency of the participant's breathing pattern. Properly configured, respiration belts have demonstrated reliable accuracy in recording respiration rates and patterns [73, 74]. However, these devices exhibit inherent limitations, particularly their inability to distinguish between thoracic and abdominal breathing. This limitation can result in inaccurate recordings when participants alter their breathing style, or if they have specific health conditions such as obesity or diaphragmatic paralysis that affect breathing mechanics. Moreover, technical issues like improper belt attachment to the transducer, inadequate fit, or compromised connections between the belt and transducer can significantly impact the integrity of the data collected.

Addeh et al. [28] categorized the Human Connectome Project in Development (HCP‐D) respiratory belt signals into seven classes, taking into consideration the usability of the signals and spurious spike artifacts (clipping to minimum and maximum measurable values). These classes are as follows: Class 1: signals with removable spurious spike artifacts (Figure 4a–c); Class 2: signals with unremovable spurious spike artifacts (Figure 4d); Class 3: partially recorded signals (e.g., Figure 4e); Class 4: Signals with very low amplitude (Figure 4f); Class 5: cases not recorded (Figure 4f); Class 6: very distorted signals with high‐frequency noise (Figure 4g); and Class 7: connections changed at a certain point (Figure 4h). It is noteworthy that all respiratory time courses contain some degree of spurious artifacts; while some can be corrected through signal processing, others are too severe or are distorted by high‐frequency noise, rendering them unusable. Notably, in our previous work using large public databases, a significant portion of the respiratory signals—up to 87%—fall into Classes 2 to 7, rendering them unusable due to severe spurious artifacts, technical malfunctions, subject movement, or improper belt‐positioning, which often leads to sudden changes in signal amplitude or incomplete data capture [28].

FIGURE 4.

Example of respiratory signals from different groups in the human connectome project development [75]. (a–c) Signal with removable spikes (HCP‐D‐0425335, Session 1, Run 2; HCD0271031, Session 2, Run 2; HCP‐D‐2000111, Session 2, Run 1). Light gray graphs show the original respiratory signal with removable spikes. (d) Signal with unremovable spikes marked by color dashes (HCP‐D‐2335344, Session 2, Run 2); (e) partially recorded signal (HCP‐D‐0968878, Session 2, Run 2). In this example, spikes are removed from recorded parts of the signal to show the impact of spike elimination clearly; (f) Not recorded (signal with zero amplitude, HCP‐D‐0110411, Session 2, Run 1) and signal varies only in a small range having square pulse‐shape pattern (HCP‐D‐0694564, Session 1, Run 2); (g) very distorted signals with high‐frequency noise (HCP‐D‐0146937, Session 1, Run 1); (h) connections changed at a certain point (HCP‐D‐1796577, Session 1, Run 1). Adapted from Addeh et al. [28].

When the subject does not comply, moves, or talks during the experiment, it introduces noise and artifacts into the data, which can reduce the statistical power, making it harder to detect real effects. A synchronized video recording of the subject can help in the identification of artifacts like subject movement or talking. The experimenter could also manually enter markers in the data during the experiment whenever the subject moves, talks, and so on. The second approach is ideal if the respiration cycles can still be visually identified, but it relies on human judgment to identify respiration cycles and introduces subjectivity.

4.2. Cardiac Activity

Cardiac activity monitoring during fMRI is essential for distinguishing between neurologically relevant signals and physiological noise caused by cardiovascular variations. Pulse oximetry, which utilizes photoplethysmography, is a primary method for monitoring cardiac activity. This technique relies on light absorption differences in blood‐infused tissues modulated by the cardiac cycle to assess global blood oxygenation levels in tissue [76]. Its application spans various fMRI protocols, from task‐based to resting‐state studies.

Despite its widespread use, the accuracy of pulse oximetry can be compromised by several factors. Ambient lighting can interfere with the optical components of the device, and physiological characteristics such as skin pigmentation or the presence of peripheral vascular disease can lead to imprecise measurements. Additionally, motion artifacts—such as the finger movements—can introduce significant errors in the recorded data, challenging the reliability of the results.

To circumvent these issues, alternative systems such as electrocardiograms (ECG) and vectorcardiograms (VCG) can be employed. These modalities are less affected by external lighting and pigmentation variability, providing more consistent and accurate recordings across diverse populations. ECG captures a planar (2D) projection of cardiac electrical activity, while VCG offers a three‐dimensional (3D) spatial representation, allowing for greater detail and precision in cardiovascular monitoring during fMRI. However, setup complexity, participant comfort, and susceptibility to artifacts often make pulse oximetry the preferred choice. For example, gradient‐induced artifacts in ECG signals, caused by the MRI scanner's imaging gradients, can introduce substantial signal distortions that overshadow intrinsic cardiac activity. Standard filtering methods used to address these artifacts may inadvertently remove critical signal details, reducing ECG's utility to basic heart rate detection, which pulse oximetry can achieve more effectively. Balancing these trade‐offs is essential for selecting the most suitable cardiac monitoring method in fMRI studies, ensuring both data fidelity and participant feasibility.

4.3. P_ETCO₂ Level

Capnography utilizes infrared spectroscopy to determine CO₂ levels of the gas at the end of exhalation, reflecting alveolar gas concentration and indirectly marking arterial CO₂. While P_ETCO₂ measurements have a more direct correlation with the mechanisms driving BOLD signal changes compared to the variations in respiration rate and depth measured by a respiratory belt ($\mathrm{\Delta }\mathit{\text{Respiration}}⟶{\mathrm{\Delta }\mathit{CO}}_2⟶∆S$), their use in fMRI settings is less common than respiratory belts because of several reasons.

First, unlike continuous respiratory signals obtained from respiratory belts, P_ETCO₂ data is collected only during exhalation. This results in intermittent data points rather than a continuous stream, which can complicate the modeling and correction of respiratory‐induced variations in BOLD signals. The reliance on breath‐by‐breath data means that the temporal resolution of P_ETCO₂ measurements is inherently lower compared to continuous respiratory monitoring methods.

Second, the implementation of P_ETCO₂ monitoring necessitates the use of either a nasal cannula or a mouthpiece, both of which may be perceived as intrusive and uncomfortable by participants. For instance, mouthpiece systems, often resembling snorkels, can exacerbate feelings of claustrophobia, particularly in individuals undergoing their first fMRI scan, thereby reducing participant tolerance and compliance. Similarly, nasal cannulas, while less obtrusive, can still induce significant discomfort and may be equally aversive for some subjects. These challenges are further amplified in studies involving prolonged scanning sessions or tasks that disrupt natural breathing patterns. Consequently, participant tolerance should be prioritized when selecting a P_ETCO₂ monitoring system, as discomfort not only influences participant compliance but also compromises data quality and overall study outcomes.

Third, the accuracy and reliability of P_ETCO₂ measurements are highly dependent on consistent and proper breathing behavior from subjects. Variations in how subjects breathe, especially inconsistencies in nasal breathing during nasal cannula sampling, can introduce noise and reduce the quality of P_ETCO₂ data. Subjects are often instructed to breathe through their nose to ensure accurate CO₂ sampling; however, adherence to this instruction can be variable (especially children, elderly, and disease population), leading to potential inaccuracies in the recorded data.

Fourth, P_ETCO₂ monitoring devices, such as capnographs, may periodically perform self‐calibration processes that last for several seconds. During these periods, CO₂ data is not logged, resulting in gaps that can disrupt the continuity of the recorded physiological data. These interruptions necessitate either the exclusion of affected data segments or interpolation, both of which can impact the overall data quality and the effectiveness of physiological noise correction [50].

Fifth, there is also a time lag between the occurrence of a breathing cycle and the corresponding change in P_ETCO₂ that reaches the capnograph. This delay, which can be several seconds, must be accounted for when correlating P_ETCO₂ data with BOLD signal changes. The latency can vary across individuals and even within different regions of the brain, complicating the interpretation and correction of P_ETCO₂‐related physiological confound [50].

In summary, while P_ETCO₂ monitoring is a powerful tool for assessing respiratory confounds in fMRI, its intermittent nature, susceptibility to noise, self‐calibration interruptions, delayed response, and practical limitations hinder its widespread adoption. Even large‐scale initiatives like the HCP and the Adolescent Brain Cognitive Development (ABCD) study, which each include more than 5000 subjects, did not employ capnography; instead, they opted to monitor respiration using respiratory belts.

5. Performance of Physiological Measures

5.1. Respiratory Measures

Employing respiratory belt measurements as a proxy for P_ETCO₂ is markedly effective in removing low‐frequency respiratory confounds in fMRI data analysis. While the respiratory belt lacks the capability to measure airflow quantitatively—a strength of capnography—innovative semi‐quantitative indices like respiratory volume per time (RVT) and respiratory variation (RV) have been developed to capture respiratory dynamics from belt data. RVT quantifies the respiratory cycle's amplitude normalized by duration, while RV measures the variability of the respiratory signal within a brief temporal window. Birn et al. [11] introduced RVT as a regressor and Chang et al.'s [14] subsequent application of RV in the General Linear Model (GLM) have significantly enhanced the mapping of neural activation and connectivity patterns, offering a robust mitigation of respiration‐related signal perturbations. Figure 5a showcases the ability of RVT and RV to track changes in the depth and rate of respiration using data from the HCP‐D dataset.

FIGURE 5.

Comparative analysis of respiratory measures in fMRI data correction. Panel (a) displays the synchronization of RVT and RV with respiratory depth and rate using HCP‐D dataset. Panel (b) highlights RVT and RV's sensitivity to a deep breath event (green arrow) and a missed deep breath by RVT (blue arrow). Panel (c) demonstrates the disparate responses of RVT and RV to simultaneous changes in breathing rate and depth (orange double arrows). Panel (d) illustrates an 11‐s breath‐holding period's impact on RVT and RV (orange arrow), with RV showing an initial decline followed by stabilization and increase, and RVT's variable response due to adjacent time series values.

While RVT and RV effectively quantify breathing dynamics, they do not capture the temporal influence of these dynamics on fMRI signal variation. Birn et al. [23] addressed this by introducing the Respiration Response Function (RRF), which models the timing of fMRI signal shifts due to respiration. By convolving RVT with RRF, this method enables precise prediction of signal changes. Extending this method, RV is also convolved with the RRF introduced by Birn et al. [23], in subsequent studies [14, 50, 77], enhancing the correction of respiratory artifacts in fMRI analysis. In essence, measures like RVT and RV serve to index respiratory “events,” while customized hemodynamic response functions relate to the ensuing BOLD signal changes. Therefore, it is crucial that these measures capture all respiratory events accurately, and that the corresponding hemodynamic response functions accurately depict the BOLD signal modifications. Deep breaths and breath‐holding are crucial respiratory events that significantly impact the BOLD signal—deep breaths typically decrease it, while breath‐holding tends to increase it. These effects are reflected in the shapes of RVT and RV, as a deep breath leads to an increase and breath‐holding results in a decrease. Our previous work [78] compared these measures due to their substantial influence on the BOLD signal. Figure 5b illustrates a deep breath detected by both RVT and RV (indicated by the green arrow). Nevertheless, the performance of RVT and RV raises several issues. Notably, RVT may not detect deep breaths as easily as RV, exemplified in Figure 5b (blue arrow) where a deep breath, longer than usual, was missed by RVT but identified by RV. Recent advancements address these limitations. For instance, Harrison et al. [79] proposed a Hilbert‐transform‐based method to estimate RVT, bypassing the need for peak detection. This method processes respiratory signals by decomposing them into instantaneous amplitude and phase components, with amplitude reflecting breathing depth and phase capturing breathing rate. Unlike traditional peak‐based estimators, this approach enables time‐resolved RVT estimation by leveraging the continuous nature of the Hilbert transform.

Differing computation methods for RVT and RV from respiratory traces mean they respond distinctly to changes in depth and rate. RVT correlates strongly with both depth and rate, while RV correlates mainly with depth. This differential response is depicted in Figure 5b,c; in Figure 5b (orange arrow), RVT identifies a rapid breathing rate not detected by RV, and in Figure 4c, marked by orange double arrows, a decrease in RVT and an increase in RV are observed. Both indices employ the RRF, so changes in RV reduce BOLD signal amplitude, whereas RVT does the opposite. A pattern of slow, deep breathing, as seen where the subject breathes deeply at a reduced rate, can neutralize CO₂ changes, posing questions about which factor affects BOLD signals more. This discrepancy and the need for further research, especially with P_ETCO₂ levels, are evident.

Lastly, RV and RVT display variable behaviors during breath‐holding. Figure 5d illustrates an approximate 11‐s breath‐holding phase (indicated by the orange arrow), affecting both metrics. RV initially decreases, stabilizes, and subsequently increases, a pattern possibly not observed for breath‐holds shorter than 4 s when using a 6‐s window for RV, which may change with different window sizes for RV calculation. When convolved with the RRF, these indices influence BOLD signals in different directions; typically, the BOLD signal is expected to only increase during breath‐holding. Most crucially, breath‐holding following expiration leads to a rapid and more notable increase in CBF and BOLD signal compared to breath‐holding after inhalation. Specifically, post‐expiration breath‐holding triggers an immediate surge in the BOLD signal, while the BOLD signal initially diminishes during post‐inhalation breath‐holding [45]. RV does not account for this discrepancy. As for RVT, its trend during breath‐holding is determined by the values of the preceding and subsequent time series, resulting in unstable performance across different respiratory patterns.

5.2. Heart Rate Variation

Chang et al. [14] introduced a model accounting for the influence of low‐frequency heart rate fluctuations on the BOLD signal. This model calculates heart rate variations (HRV) as the average time interval between consecutive R peaks within a six‐second window, offering a direct measure of the cardiac influence on the BOLD signal. Concurrently, they proposed the cardiac response function (CRF), which delineates the temporal dynamics of the BOLD signal in relation to HR variations [14]. The convolution of HRV time series with the CRF enhances the GLM's capacity to explain the variance in the BOLD signal due to cardiac cycles. Unlike respiratory measures such as RVT and RV, which may not detect all respiratory events or display variable performance during events like breath‐holding, the HR variation provides a stable and reliable index of cardiac‐related BOLD signal fluctuations, attributable to the intrinsic control of the autonomic nervous system over the heart rate, which is not subject to voluntary manipulation.

However, the accuracy of HR measurement can be affected by motion artifacts and noise in the PPG or ECG signals used to detect R peaks. These artifacts can introduce errors in HR estimation and subsequently in the modeled HR influence on the BOLD signal. Motion artifacts, in particular, can distort the recorded cardiac signal, leading to inaccurate identification of R peaks. Noise in the signal can also mask true HR variations or create false fluctuations, further complicating the accurate modeling of HR‐related BOLD signal changes. Additionally, the effect of HR on the BOLD signal varies across different brain regions, as demonstrated by the variability in the CRF [14]. As a result, ensuring high‐quality, artifact‐free HR data is crucial for reliable correction of physiological noise in fMRI studies.

6. Correction Methods

This section outlines the diverse strategies employed to mitigate the effects of physiological confounds, categorized into three main types: filtering methods, model‐based methods, and data‐driven methods. Each strategy is tailored to address different aspects and characteristics of the confounds they intend to correct.

6.1. Filtering Methods

In a straightforward approach to addressing physiological confounds within BOLD signals, primary frequencies of respiratory and cardiac fluctuations are first identified from peaks in the spectrum of external recordings. Notch filters are subsequently applied directly to these frequencies to selectively reduce high‐frequency respiratory and cardiac confounds [22]. When the TR is sufficiently short, these primary frequencies do not alias with those associated with the BOLD signal, allowing for their effective removal without affecting the BOLD response, which typically occurs below 0.1 Hz. To prevent the aliasing of high‐frequency respiratory and cardiac confounds into lower frequencies, which typically appear around 0.3 and 1 Hz, respectively, adherence to the Nyquist criterion necessitates a TR of less than 1.6 s for respiratory rates and less than 0.5 s for heart rates. If the TR does not meet these criteria, physiological data (e.g., from a pulse oximeter) are resampled at the same rate as the fMRI acquisition to precisely determine how physiological frequencies are aliased due to the suboptimal TR. Upon identifying the aliased frequencies, tailored notch filters are applied to the fMRI data. The design of these notch filters specifically targets the aliased frequencies of cardiac and respiratory rates, effectively mitigating their impact on the fMRI signal. This method, as reported by Biswal et al. [22], has been effective in reducing signal variability and enhancing the detection of stimulus‐related signal changes, even when the TR is not sufficiently short to accurately sample the primary frequency.

However, notch filtering presents significant challenges, particularly its potential to inadvertently eliminate BOLD signal fluctuations that coincide with physiological frequencies targeted for removal. For example, with a standard TR of 2 s in whole‐brain imaging, respiratory and cardiac frequencies—such as those from a child with a respiratory rate of 26 breaths per minute (0.43 Hz) and a heart rate of 85 beats per minute (1.41 Hz)—will alias to 0.067 and 0.083 Hz, respectively. These frequencies fall within the range typically associated with BOLD signal fluctuations. Consequently, employing a notch filter in such scenarios can inadvertently suppress crucial BOLD signal that are integral for identifying neurophysiological responses at these frequencies. Furthermore, physiological fluctuations exhibit considerable variability throughout an fMRI scan, notably in task‐based studies, but also in resting‐state conditions. For instance, variations in respiratory rates that exceed 0.2 Hz during an fMRI scan [80], lead to significant shifts within the respiratory frequency spectrum, introducing multiple peaks. Given these dynamics, there is a critical need to design notch filters that are adaptive to such variations. However, this approach may lead to the further elimination of neural‐related BOLD signal fluctuations that coincide with these physiological confounds.

Traditional notch filtering approaches often fail to consider the power spectrum of task‐related activity in their design, which may compromise effectiveness and result in the loss of crucial task‐specific neural signals. This issue is particularly relevant in event‐related designs with short intertrial intervals, or in block‐ and event‐related designs where the TR exceeds 0.5 s [81]. Although it is possible to repeat the same task at different rates to avoid overlap, this practice can introduce fatigue or habituation effects in participants, potentially altering physiological and neural responses and complicating the interpretation of results. Alternatively, the adaptive Wiener filtering approach proposed by Buonocore and Maddock [82] offers a refined method that derives confound characteristics directly from the data. This technique enhances signal fidelity by estimating the power spectrum of the physiological confound from ventricular voxels, random confound from white matter voxels, and task‐related activity directly from the fMRI signals. However, this method assumes uniform spectral characteristics of physiological fluctuations across the entire brain, which may not accurately reflect the complex variability observed in different brain regions. Such assumptions necessitate a cautious interpretation of the filtered data, as they can significantly influence the reliability of the outcomes.

Besides notch filtering, bandpass filtering is a prevalent technique in resting‐state fMRI studies. This method typically excludes frequencies below 0.01 Hz and above 0.1 Hz, ensuring that fluctuations outside the frequency range of interest do not impact functional connectivity measures [16]. While bandpass filtering effectively addresses high‐frequency respiratory and cardiac confounds, its efficacy is contingent on a sufficiently short TR (< 0.5 s). At longer, more standard TRs (e.g., 2 s), these confounding frequencies are aliased into the lower frequency range [83]. Additionally, frequency filtering strategies are limited in their ability to remove low‐frequency respiratory confounds (~0.03 Hz) and low‐frequency cardiac confounds (~0.04 Hz), which overlap with the frequency range of neural activity‐related resting‐state BOLD oscillations [16].

The removal of genuine neural signals due to aliasing in both task‐based and resting‐state fMRI studies, along with the inability to effectively remove BOLD signal components induced by low‐frequency respiratory and cardiac fluctuations, limits the effectiveness of filter‐based approaches. Consequently, more sophisticated model‐based and data‐driven methods have been developed to effectively reduce physiological confounds.

6.2. Model‐Based Methods

Model‐based confound removal methods utilize additional physiological recordings, such as those of the respiration and heartbeat, which are now relatively straightforward to obtain on many modern scanners.

6.2.1. Addressing High‐Frequency Confounds

Model‐based methods to remove the high‐frequency respiratory and cardiac confounds have been developed for both k‐space [84, 85, 86, 87], and image space data [18]. The most widely used model‐based correction method in the frequency domain is retrospective k‐space correction, referred to as RetroKCor [85]. RetroKCor involves analyzing the fMRI dataset alongside synchronized respiratory and cardiac data to detect and model the patterns of these physiological activities, quantifying their impact on each point in the k‐space data. The method then applies corrections tailored to the specific temporal and spatial characteristics of the physiological confound. These corrections are made directly to the k‐space data, which are subsequently used to reconstruct fMRI images with significantly reduced physiological confound, thereby enhancing both the accuracy of the imaging data. Several updates have been proposed to enhance the performance of RetroKCor. For instance, the method developed by Le and Hu [86] improves confound modeling by extracting physiological parameters directly from the MR data, thus eliminating the need for external monitoring. Additionally, the method developed by Frank et al. [87] optimizes spatial frequency handling by utilizing unaliased confound information from multislice data, leading to better estimation and reduction of high‐frequency respiratory fluctuations.

However, RetroKCor is subject to several limitations that can compromise its effectiveness. This method's correction of individual k‐space points impacts all voxels, potentially introducing spatially correlated confound across the image domain. RetroKCor also permits high spatial frequency confounds to persist, as the method's inherent limitations in modeling and subtracting complex confound patterns at finer spatial scales may leave these confounds unaddressed. Moreover, the accuracy of fitting Fourier terms varies significantly, being more precise in the inner frequencies of k‐space compared to the outer frequencies due to variable signal‐to‐noise ratio (SNR) [27]. This discrepancy can lead to spatial blurring following the transformation of k‐space data to the image domain. Furthermore, the corrections may inadvertently modify the signal in areas with minimal physiological artifacts, potentially obscuring genuine neural activity. Moreover, RetroKCor struggles at high field strengths like 7 Tesla due to its inability to adequately address increased physiological confound and susceptibility artifacts, originally optimized for lower fields [88]. The method's limitations in modeling complex physiological fluctuations and the need for precise synchronization of k‐space data acquisition at higher resolutions make it less effective without adaptations tailored to high‐field MRI environments.

RETROspective Image CORrection (RetroICor) [18] is a widely recognized method for mitigating high‐frequency respiratory and cardiac confounds in fMRI data. The concept was originally proposed by Josephs et al. [89], who utilized Statistical Parametric Mapping (SPM) to account for physiological motion components as nuisance regressors. Their approach integrated ECG and respiratory phase waveforms into statistical models as nuisance regressors, thereby allowing for the partial attenuation of physiological confound in fMRI analyses. However, this technique was primarily designed for statistical modeling of physiological fluctuations rather than direct retrospective correction of image time series, limiting its utility in applications requiring explicit physiological artifact removal. Recognizing the need for a more generalizable and independent correction method, Glover et al. [18] developed what is now widely known as RetroICor.

RetroICor assumes that respiratory and cardiac fluctuations can be modeled as quasi‐periodic processes. RetroICor employs a low‐order Fourier series, specifically of order 2, to model these physiological fluctuations, fitting time‐varying respiratory and cardiac phases to the data as nuisance regressors, which are subsequently removed. The respiratory phase is estimated using a histogram‐equalized method derived from the amplitude of the respiratory signal, while the cardiac phase is determined based on the interval between the time of image acquisition and the previous cardiac peak, relative to the R‐peak interval. By synchronizing the phases of the respiratory and cardiac cycles with the timing of each imaging slice, RetroICor effectively addresses aliasing in low frequencies and manages non‐stationary respiratory and cardiac frequencies. Since RetroICor estimates the respiratory phases using a histogram‐based approach that considers the entire waveform of the signal, rather than using a peak detection algorithm, it effectively removes respiratory‐induced signal changes. This is because susceptibility changes and head movements depend on the entire shape of the breathing cycle, not just the timing of peak inspiration. The respiratory phase needs to increase nonlinearly over time due to significant changes around the times of inspiration and expiration, remaining constant in between. A peak detection approach can only model a linear increase in the respiratory phase, which might not be sufficient except for short TR acquisitions and if comprehensive modeling of the motion‐related effects is performed.

In contrast to k‐space methods, RetroICor is implemented in the image domain, thus circumventing several constraints associated with k‐space techniques. Additionally, given that most fMRI studies do not readily have access to k‐space data, the RetroICor method has become the more practical choice in widespread use. However, a critical limitation of RetroICor is its assumption that the Fourier coefficients, representing physiological fluctuations, remain constant over time. This assumption of static amplitudes is simplistic and does not hold in practical scenarios, where physiological dynamics exhibit temporal variability. This discrepancy between the model's assumptions and the dynamic nature of physiological signals can significantly impact the accuracy of the correction process [90]. Additionally, RetroICor considers only independent cardiac and respiratory fluctuations, which may lead to incomplete correction of physiological confound. This limitation can particularly affect the accuracy of fMRI data in regions like the brainstem, where interactions between cardiac and respiratory signals are significant. Harvey et al. [91] address this issue by including multiplicative terms that account for these interactions. This modified approach fits a Fourier series with three cardiac harmonics, four respiratory harmonics, and one interaction term (denoted as the “3C4R1X” model), significantly reducing signal variability without overfitting to confound. By incorporating these interaction terms, the improved method enhances the detection of brainstem activation and increases the overall reliability of fMRI data. Additionally, while the original RetroICor at 1.5 T showed minimal benefits from increasing the model order above 2, investigations at 3 T have demonstrated improved model fitting with higher orders [91], albeit with the trade‐off of potentially losing the signal of interest due to increased complexity.

Despite its limitations, RetroICor remains a more effective technique for correcting high‐frequency physiological confound in fMRI data than k‐space correction methods such as RetroKCor and its modified versions [88].

6.2.2. Addressing Low‐Frequency Confounds

Wise et al. [12] pioneered the exploration of CO₂‐induced BOLD signal fluctuations in functional MRI by methodically recording P_ETCO₂ levels throughout scanning sessions. They incorporated the recorded P_ETCO₂ as a regressor in a general linear model, utilizing a gamma‐variate hemodynamic response function to account for the expected BOLD signal delay. This meticulous approach significantly mitigated the influence of CO₂ fluctuations on the BOLD signal, thereby enhancing the isolation of neural activity regions. By integrating P_ETCO₂ as a covariate, the analysis robustly reduced CO₂‐related signal variations, emphasizing the critical importance of accounting for physiological confounds in fMRI studies.

Building upon this foundation, Birn et al. [11] advanced the methodology by introducing the use of RVT to evaluate low‐frequency respiratory variations, attributing CO₂ fluctuations to subtle variations in breathing depth and rate observed during scans. This advanced technique involves correlating physiological and fMRI time series at multiple temporal shifts, thereby facilitating the identification of optimal latencies for regression analysis within the GLM framework. Subsequently, Birn et al. [23] refined this approach by introducing a specialized “Respiration Response Function” (RRF), named RVTRRF, that more accurately models the BOLD signal fluctuations induced by a variety of respiratory manipulations, such as breath‐holding and variations in breathing depth and rate. This function captures the complex temporal dynamics of signal changes associated with respiratory events more effectively than the previous model, offering a more nuanced analysis of respiratory influences on the BOLD signal. The name RVTRRF reflects the convolution of RVT with RRF, signifying its role in modeling the BOLD response to respiratory variations more comprehensively than RVT alone.

A study by Shmueli et al. [13], as discussed previously in Section 3.2, demonstrated that temporal correlations between heart rate and BOLD signal can reveal substantial additional variance when included as regressors in the analysis, thereby refining the confound reduction techniques in fMRI studies. They employed time‐shifted heart rate data as regressors alongside established physiological confound models such as RETROICOR and RVT to enhance the accuracy of confound correction. Following this exploration, Chang et al. [14] proposed a method integrating both RV and heart rate (HR) as regressors to account for physiological variations. This method, utilizing a convolution model, further refines the accuracy of capturing BOLD signal fluctuations by addressing both respiratory and cardiac influences simultaneously. It involves the simultaneous deconvolution of HR and RV from the BOLD signal on a voxel‐wise basis, employing a linear systems approach with Gaussian process priors to generate hemodynamic response functions (HRFs). This allows the model to provide a more comprehensive accounting of physiological confound, demonstrating that including both RV and HR explains a significantly greater variance in the BOLD signal than models, including RV alone.

To further improve the accuracy of addressing low‐frequency physiological confounds in fMRI, Golestani et al. [24] have developed an integrative method that considers multiple sources of physiological confound, including CO₂, RVT, and HRV. The rationale behind considering all these factors together lies in their potential interactions and individual contributions to the BOLD signal variance. By simultaneously including CO₂, RVT, and HRV in a multi‐regression model, the study aimed to comprehensively account for physiological confound and isolate the unique impact of each factor on the fMRI signal. The study utilized a voxel‐wise estimation of Physiological Response Functions, which model the BOLD signal response to these physiological variations. Each response function was modeled using a Gaussian process, providing a flexible and robust means to capture the temporal dynamics of physiological influences on the BOLD signal. This approach allowed for a detailed characterization of the physiological confound at a regional level, enhancing the spatial specificity of the confound correction. Furthermore, Golestani et al. [51] investigated the method of clamping P_ETCO₂ and compared it with free‐breathing conditions and retroactive CO₂ correction. This technique involves maintaining a constant P_ETCO₂ level during fMRI scanning, thereby providing a stable baseline and reducing the physiological confound related to CO₂ fluctuations.

In addition to physiological recordings such as P_ETCO₂ and respiratory and cardiac signals, functional near‐infrared spectroscopy (fNIRS) presents a sophisticated model‐based approach to addressing low‐frequency physiological confounds in fMRI data [15, 92, 93, 94, 95]. By measuring changes in oxy‐hemoglobin and deoxy‐hemoglobin concentrations at high temporal resolutions (6–20 Hz), fNIRS provides an unaliased perspective on systemic low‐frequency oscillations that are challenging to capture with traditional fMRI methods due to its lower sampling rate [96]. Techniques like the Regressor Interpolation at Progressive Time Delays (RIPTiDe), introduced by Frederick et al. [92], leverage concurrent fNIRS and fMRI data to develop voxel‐specific regressors that account for temporal delays in the propagation of systemic low‐frequency oscillations. This method combines high temporal resolution of fNIRS with spatially resolved fMRI signals, enabling more precise modeling of physiological confounds. By incorporating voxel‐specific dynamics and band‐specific regressors, RIPTiDe enhances the capacity to disentangle systemic physiological signals from neural activity. Notably, RIPTiDe explains 55% more variance in the BOLD signal compared to RetroICor+RVT in key regions, such as gray matter, due to its broader physiological scope, including contributions from blood pressure changes and vasomotor fluctuations, as well as its ability to model both periodic and non‐periodic fluctuations.

The implementation of RIPTiDe is facilitated by the Rapidtide package, a Python‐based tool designed to streamline time‐delay modeling and physiological noise correction in fMRI data [97]. Similarly, the PhysIO Toolbox, available as part of the TAPAS software suite, offers robust preprocessing of physiological recordings, implements RetroICor and RVT/HRV confound models, and integrates seamlessly into Statistical Parametric Mapping (SPM) pipelines for fMRI data preprocessing and modeling [42]. These software tools exemplify the growing accessibility and sophistication of model‐based approaches in physiological confound correction.

While model‐based approaches like RetroICor and RVHR effectively mitigate high‐frequency and low‐frequency physiological confounds, their performance heavily depends on the availability and quality of physiological measurements. If these measurements are unavailable or suboptimal, the effectiveness of these methods is significantly compromised. Another critical challenge is determining the appropriate HRF to accurately model the physiological fluctuations. Unlike neuronal‐induced BOLD responses, the signal changes from physiological variations have different temporal dynamics and amplitudes, complicating the derivation of a precise HRF. This complexity requires advanced modeling techniques and extensive data, which may not always be feasible. Additionally, achieving precise synchronization between physiological signals and imaging data can be challenging, especially in high‐resolution or high‐field MRI environments where physiological fluctuations are more pronounced. These limitations highlight the need for continuous monitoring and calibration, which can be cumbersome and not always practicable in all research settings.

fNIRS, while offering high temporal resolution and a promising solution for improving physiological noise correction, exemplifies many of these broader challenges. Its implementation requires additional equipment and resources, increasing experimental costs and complexity. Moreover, the setup and calibration of fNIRS devices necessitate meticulous probe placement and subject‐specific adjustments, introducing variability in data acquisition and analysis. These logistical challenges mirror the broader issues faced by model‐based methods, such as reliance on high‐quality physiological recordings and the difficulty of integrating additional modalities into existing workflows. Consequently, while fNIRS‐based approaches like RIPTiDe demonstrate significant potential for addressing low‐frequency confounds, their adoption must carefully balance these advantages against the practical and financial constraints that often limit the scalability of advanced model‐based methods. These trade‐offs underscore the need for ongoing refinement of physiological correction techniques to enhance both accessibility and effectiveness in fMRI research.

6.3. Data‐Driven Methods

The fundamental limitation of needing external monitoring for physiological activities has spurred the development and adoption of data‐driven approaches. These methods, such as principal component analysis (PCA) and independent component analysis (ICA), operate independently of predefined models. They are capable of directly identifying and removing physiological confounds from fMRI data, offering a more flexible and potentially more effective means of handling these variations.

6.3.1. PCA‐Based Method

PCA is a statistical technique used to simplify complex datasets by transforming them into a set of uncorrelated variables called principal components (PCs). One notable application of PCA in the fMRI context is the component‐based confound correction method known as CompCor (Component‐Based Confound Correction) [25]. The CompCor method identifies significant principal components from regions of interest (ROIs) in the fMRI data that are unlikely to be influenced by neural activity, such as white matter and cerebrospinal fluid (CSF). These components are then included as nuisance regressors in the GLM used to analyze the fMRI time series. CompCor uses two primary approaches to define the confound ROIs: anatomical definition (aCompCor) and temporal standard deviation (tCompCor). The anatomical definition utilizes high‐resolution anatomical scans to delineate white matter and CSF regions, which predominantly reflect physiological confound. In standard fMRI acquisitions using 2D EPI, the mean functional image provides sufficient contrast for delineating CSF ventricles and bulk white matter regions, thereby reducing the dependence on high‐resolution anatomical data. While high‐resolution anatomical scans may still be beneficial in specific cases, such as 3D EPI at ultra‐high field (UHF) strengths, the functional mean image is generally adequate for aCompCor implementation in typical fMRI studies. The tCompCor selects voxels with high temporal standard deviation, indicating physiological confound, without requiring anatomical scans. The tCompCor approach is advantageous when anatomical data is unavailable or quick identification of confound regions is needed.

Behzadi et al. [25] demonstrated that the CompCor technique could modify the normalized temporal standard deviation in resting‐state BOLD data, with RetroICor resulting in an 8% reduction, while aCompCor and tCompCor yielded reductions of 20% and 29%, respectively. However, these reductions in temporal variance do not definitively indicate whether the decrease is due to the effective removal of physiological confounds or the unintended reduction of signal variability linked to neuronal activity. Moreover, CompCor has been adapted to handle undersampled data, such as in BOLD fMRI with longer TRs. It effectively identifies aliased physiological confound components, demonstrating its robustness across various imaging protocols and conditions.

The efficacy of PCA‐based methods in fMRI data correction heavily relies on the precise definition of confound regions‐of‐interest (nROI) from which principal components are derived. These nROIs must be accurately identified to ensure that extracted components represent confound rather than neural signals. For instance, recent studies have demonstrated detectable BOLD fMRI activations in white matter [98], challenging its previously understood role as merely a passive structure involved only in structural connectivity. Thus, in the context of PCA‐based methods, reconsidering white matter as a control region is crucial. Furthermore, regions with high temporal standard deviation (tSTD) may not solely indicate confound but could include voxels with significant stimulus‐correlated activity. This overlap can impair PCA's performance by including or excluding relevant neural signals based on their correlation with experimental conditions or stimulus‐related motion, thereby complicating the distinction between true neural activity and physiological or motion‐related confound. In scenarios involving severe motion artifacts, the extracted components might predominantly reflect motion‐related changes rather than physiological confounds. This reduces PCA's effectiveness in isolating and removing respiratory and cardiac signals, particularly when motion correlates significantly with experimental tasks. Moreover, PCA might not adequately address non‐stationarities in the data, common in longer fMRI sessions or with sporadic head movements.

Selecting the optimal number of PCs in fMRI data preprocessing is crucial for balancing confound reduction and signal preservation. One common approach is to retain enough PCs to explain a predefined percentage of the total variance, typically 50%–90%, ensuring that most of the relevant signal is preserved while minimizing confound contributions [27]. Another effective strategy is cross‐validation, where different numbers of PCs are systematically tested, and the optimal selection is determined based on performance metrics such as maximizing cross‐validated R ² or minimizing error [99]. Additionally, the “broken stick” method, as described by Behzadi et al. [25], offers a statistical approach for selecting meaningful components by comparing observed eigenvalues to those expected under a null distribution, retaining only those that exceed chance levels. Beyond variance‐based thresholds, recent studies highlight the importance of spatial patterns in PCA selection. Principal components capturing structured confound sources, such as those dominated by respiration, cardiac pulsations, or head motion, can be identified based on the spatial distribution of their corresponding PC weights. By inspecting these weights, researchers can differentiate physiological artifacts from neural signals, ensuring that components reflecting global confound sources are retained while avoiding the removal of meaningful neural activity [100]. To improve adaptability across datasets, a more flexible strategy is recommended—dynamically adjusting the number of PCs based on both explained variance and spatial patterns rather than adhering to a fixed threshold. This ensures optimal noise correction while preserving functional signals, ultimately enhancing the reliability of fMRI analyses. Although experimental data support the efficacy of PCA for periodic and block designs [25, 101], its performance may be reduced in complex event‐related fMRI experiments where the expected stimulus‐related response is not well‐defined.

6.3.2. ICA‐Based Method

ICA excels in fMRI data analysis due to its ability to identify and segregate diverse sources of signal variation, including neuronal activation, head movements, cardiac and respiratory confounds, and scanner artifacts. The method operationalizes this by decomposing fMRI data into spatial maps and their corresponding time courses, optimizing for maximum statistical independence among the components. Within the ICA framework, two main methodologies are utilized: temporal ICA (tICA) and spatial ICA (sICA). While tICA excels at handling physiological confounds, such as respiratory and cardiac fluctuations, which exhibit stronger temporal independence, sICA is more effective in removing motion‐related artifacts due to its ability to capture spatially independent components [102]. tICA enhances the statistical independence of temporal patterns, crucial for analyzing brain function dynamics, making it ideal for studies on cognitive processes or temporal responses to stimuli. In contrast, sICA focuses on spatial independence, aiding in the identification of distinct brain networks or activity patterns, particularly useful in studies of resting‐state connectivity or task‐based activations where the spatial distribution is key. Following the computation of ICA, the denoising process involves distinguishing between independent components (ICs) associated with neuronal‐related BOLD signals and those related to confound sources. Subsequently, confound‐related ICs are removed, allowing for the accurate reconstruction of the dataset.

In practice, manual classification of the ICs is very time‐consuming, difficult to reproduce, and requires expertise [26]. Therefore, several procedures have been proposed for assisting automated classification, which mainly differ in the algorithms used for supervised classification (and if necessary feature selection), the number and definitions of the spatial and temporal features used in the classification, as well as the type of fMRI data they are optimized to work with, either task‐based, resting state, or both [26, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112]. A notable implementation in this area is FIX (“FMRIB's ICA‐based X‐noiseifier”), which provides an automatic solution for denoising fMRI data via accurate classification of ICA components [26]. For each ICA component, FIX generates a large number of distinct spatial and temporal features, each describing a different aspect of the data. The set of features is then fed into a multi‐level classifier. Once trained through the manual of a sufficient number of training datasets, the classifier can then automatically classify new datasets. On conventional resting‐state fMRI (rfMRI) single‐run datasets, FIX achieved about 95% overall accuracy. On high‐quality rfMRI data from the Human Connectome Project, FIX achieves over 99% classification accuracy, and as a result, is being used in the default rfMRI processing pipeline for generating HCP connectomes. FIX is publicly available as a plugin for FSL.

While ICA‐based methods are powerful tools for denoising fMRI data, their accuracy heavily depends on the correct classification of components as “signal” or “confound.” Misclassifications can lead to the erroneous removal of neural signals or the retention of confound, significantly impacting data quality. Additionally, the effectiveness of ICA is contingent upon the quality and quantity of data, with poor data quality or insufficient sample sizes potentially leading to incomplete confound identification. A critical challenge arises from the assumption of statistical independence among signals, as many brain processes are correlated, separation of confound from true neural activity is complicated. Furthermore, the manual classification and interpretation of ICA outputs require substantial expertise in fMRI technology, presenting a barrier to effective application. The process also necessitates careful consideration of the number of components to retain. If too many components are retained, confounding noise might be mistakenly included in the analysis, whereas retaining too few components risks excluding significant neural signals that are crucial for understanding brain activity. Additionally, ICA's performance can vary significantly across different fMRI conditions, such as task‐based versus resting‐state setups, necessitating study‐specific adjustments and validations. These factors underscore the need for careful application and ongoing validation of ICA‐based methods within the specific contexts of individual research projects.

6.4. Summary of Correction Methods

Physiological confound correction methods in fMRI can be categorized into filtering, model‐based, and data‐driven approaches, each with distinct advantages and limitations. Filtering methods, such as notch and bandpass filters, are simple and effective in targeting specific frequency ranges but may inadvertently remove neural‐related BOLD signals, particularly when confounds overlap with neural frequencies or when low‐frequency fluctuations are present. Model‐based methods rely on physiological recordings to accurately model and correct confounds, offering high precision but facing challenges such as assumptions about static signal characteristics and dependency on data quality. Advanced techniques like fNIRS‐based methods enhance temporal resolution but require complex setups and additional hardware. In contrast, data‐driven methods, such as PCA and ICA, operate without external recordings, providing flexibility and automation but presenting challenges in component selection, which may lead to the removal of neural signals or incomplete correction. These approaches highlight the trade‐offs between simplicity, precision, and feasibility, with Table 1 summarizing their key characteristics and associated tools for implementation.

TABLE 1.

Summary of physiological confound correction methods and tools in fMRI.

Method type	Method name	Advantages	Disadvantages	Tools/software
Filtering	Notch filtering [22]	Simple to implement; effectively removes targeted frequencies	May remove neural‐related BOLD fluctuations due to aliasing; cannot address low‐frequency confounds	Custom MATLAB and Python scripts
Filtering	Bandpass filtering [22]	Helps retain relevant BOLD frequencies	Ineffective for aliased physiological noise; does not address low‐frequency fluctuations	Custom MATLAB and Python scripts, 3dBandpass (AFNI), fslmaths ‐bptf (FSL)
Model‐based	RetroICor [18]	Directly models and removes cardiac and respiratory signals; widely used	Assumes static Fourier coefficients; does not capture dynamic physiological variations	3dretroicor (AFNI), PhysIO (SPM)
Model‐based	RetroKCor [85]	Corrects artifacts at k‐space level before image reconstruction	Requires access to raw k‐space data, which is not always available	Custom MATLAB, Python, and C++ scripts
Model‐based	Physiological modeling and regression methods [11, 12, 13, 14, 23, 24, 42, 51]	Accounts for delayed physiological responses	Relies on external physiological recordings which may be noisy or missing	PhysIO (SPM), glm (FSL), custom MATLAB scripts, 3dDeconvolve (AFNI), Nilearn GLM (Python)
Model‐based	fNIRS‐based methods (e.g., RIPTiDe) [92]	High temporal resolution; captures systemic fluctuations	Requires additional hardware and careful setup	Rapidtide (Python), custom implementations
Data‐driven	PCA [25]	Does not require physiological recordings; can be automated	Choosing the noise region and selecting the correct number of components is challenging; may remove neural signals	CONN Toolbox (SPM), custom MATLAB scripts, 3dPCA (AFNI), Nilearn (Python)
Data‐driven	ICA [26]	Effectively separates neural and non‐neural signals	Requires manual or automated classification; computationally intensive	MELODIC (FSL), ICA‐AROMA (Python), FIX (FSL), CONN Toolbox (SPM), 3dICA (AFNI), GIFT Toolbox (MATLAB), Nilearn (Python)

7. Physiological Confounds and Motion Correction

Physiological confounds affect BOLD‐fMRI data in various ways, including the estimation of head motion parameters. In earlier fMRI research with longer repetition times (TRs), the correlation between physiological processes and head motion estimates was not observable. However, technological advancements, such as multi‐band acceleration, now enable more rapid imaging with sub‐second temporal resolutions. This higher temporal resolution provides a detailed, yet seemingly noisier, view of estimated head motion traces (e.g., TRs of 0.72 s in the Human Connectome Project [113, 114]), contrasting sharply with their seemingly static appearance seen with multi‐second TRs in previous studies (e.g., a TR of 3 s in the Brain Genomics Superstruct Project [115]). fMRI researchers have noted that the increased variability in estimated head motion traces observed in fast‐TR data arises, in part, because certain physiological cycles, especially respiratory ones, no longer alias to lower frequencies, making their effects more visible in motion estimates [32]. This advancement in TRs has fundamentally shifted our understanding and the methods used for motion correction in fMRI data.

7.1. Respiration and Motion Correction

During fMRI, respiratory activities contribute to two distinct types of head motion: actual movement and pseudomotion. Actual movement typically manifests as head nodding, primarily along the vertical position and pitch axis, and can have a substantial magnitude. In contrast, pseudomotion arises due to respiration‐induced B₀ field fluctuations, which can mimic motion in the phase encoding (PE) direction—even when no actual head displacement occurs. In single‐shot EPI, these B₀ fluctuations cause global off‐resonance effects, leading to positional voxel shifts predominantly in the PE direction [20]. The corresponding resonance frequency shifts result in spatial misregistrations, where the voxel displacement $∆y$ in the PE direction is given by

$$\frac{∆y}{p}\approx ∆f\times T_{\mathit{acq}}$$ (3)

where $p$ is the pixel size, $∆f$ is the frequency shift due to off‐resonance effects, and $T_{\mathit{acq}}$ is the total k‐space readout duration [116]. It is worth noting that $1/T_{\mathit{acq}}$ corresponds to the pixel bandwidth—a parameter that some scanner vendors report in place of the readout duration. This phenomenon becomes more pronounced at higher magnetic fields (e.g., 7 T MRI) due to the increased sensitivity of the Larmor frequency to even small variations in B₀.

This artifact presents a challenge to retrospective image realignment algorithms, which may mistakenly classify these artifacts as real head motion. Furthermore, modulations of the B₀ field can extend to other directions, rendering pseudomotion detectable across all head motion parameters [32, 117], which introduces a frequency component at approximately 0.3 Hz, corresponding to the breathing rate, into all estimated motion metrics. Addeh et al. [80] explored this phenomenon across diverse age cohorts, recognizing that respiratory rates inherently vary with age. This investigation utilized three datasets: HCP‐D, HCP‐YA, and HCP in Aging (HCP‐A), where HCP‐A participants ranged from 36 to over 100 years old [113]. The primary respiratory frequencies extracted from the HCP‐D, HCP‐YA, and HCP‐A cohorts were $0.324\pm 0.083$, $0.303\pm 0.051$, and $0.272\pm 0.093$ Hz respectively, paralleling the head motion frequencies of $0.318\pm 0.068$, $0.293\pm 0.063$, and $0.263\pm 0.071$ Hz. These values correspond to the conventional respiratory rates observed within each age group. A paired t test revealed no significant difference between the respiratory frequencies and corresponding head motion frequencies (p > 0.01) [80]. Figure 6 illustrates the power spectral density profiles of motion traces from the HCP‐D, HCP‐YA, and HCP‐A datasets, highlighting the influence of respiration on estimated head motions during fMRI sessions. Oscillations around 0.3 Hz are particularly noticeable in the phase‐encode direction, attributable to both actual head movement due to respiration and the pseudomotion artifact. Additionally, the presence of peaks at approximately 0.12 Hz is associated with sporadic deep breathing events. Importantly, these oscillations in the estimated head motions at breathing rates can be advantageous. For example, Addeh et al. [118] leveraged estimated head motions combined with BOLD signals as inputs for a deep learning model that reconstructs the respiration variation signal. This indicates that estimated head motions can provide informative data about the breathing rate of the subject and assist with correcting both low‐frequency and high‐frequency respiratory confounds in fMRI data.

FIGURE 6.

Power spectral density of motion traces for HCP‐D dataset, HCP‐YA dataset, and HCP‐A dataset. In HCP‐D and HCP‐A project, the PE direction is anterior → posterior (AP) and posterior → anterior (PA). In HCP‐YA project, the PE direction is $\mathit{\text{Left}}\to \mathit{\text{Right}}$ (LR) and $\mathit{\text{Right}}\to \mathit{\text{Left}}$ (RL). The respiration creates head pseudomotion at frequency of ~0.3 Hz as shown by red arrow, which consistent with the normal breathing rate of each group. The low‐frequency motion peak, 0.12 Hz, was not evident in cohort of children but was present in all adult cohorts, especially in the aged population as shown by blue arrow.

In fMRI research, data censoring is a critical step where time points with excessive motion are selectively excluded to minimize artifactual influences on functional connectivity analyses. This process requires careful application to avoid discarding informative data inadvertently. Recent methods developed by Power et al. [32], refined by Fair et al. [117], and further explored by Kaplan et al. [119], recommend using notch filters to attenuate respiratory‐induced motion artifacts. These filters have significantly improved data quality and enhance the robustness of functional connectivity metrics by preserving more analyzable data. Notably, Kaplan et al. [119] demonstrated that applying these filters in the Baby Connectome Project [120] and the Early Life Adversity, Biological Embedding datasets increased the mean usable data duration from $1.8\pm 2.5$ to $13.9\pm 5.2$ and from $9.6\pm 5.3$ to $19.4\pm 5.2$ min, respectively. This highlights the importance of tailored data preprocessing in enhancing the utility of fMRI datasets for neuroscientific research.

Nevertheless, using fixed‐frequency band notch filters, as implemented in prior studies [32, 117, 119], presents notable limitations. First, such filters fail to accommodate the variability in respiratory rates observed during fMRI sessions, which can sometimes fall significantly below the filter's range, as seen with periods under 0.1 Hz in the Human Connectome Project datasets. An instance of this is depicted in Figure 7, where respiratory rates across various datasets demonstrate extensive variability, at times plunging below 0.2 Hz. A notch filter preset to a particular frequency range is effective in reducing head motion amplitudes near the average respiratory rate (approximately 0.3 Hz) but falls short in capturing slower respiration‐induced movements. In the illustrated case (Figure 7), the notch filter's cut‐off frequency is set at 0.35 Hz with a bandwidth of 0.3 Hz, following Power et al.'s protocol [32]. Second, while notch filters are adept at mitigating pseudomotion, their specificity may also lead to the unintended removal of genuine respiratory‐related head movements, such as nodding, since both types of motion manifest at similar respiratory frequencies.

FIGURE 7.

Comparison of respiratory signals and motion parameters in fMRI data. The graphs illustrate the raw respiratory signal (black), unfiltered motion parameters in the PE direction (red), and motion parameters post‐notch filter application (blue). Notably, the notch filter, with a cut‐off frequency set at 0.35 Hz and a bandwidth of 0.30 Hz, reduces respiratory‐related motions. However, it does not capture the full spectrum of respiratory rate variability, which includes periods of slower breathing rates dipping below 0.2 Hz.

Addressing these challenges requires more advanced correction strategies that leverage field monitoring and signal processing techniques. NMR field probes provide high‐precision dynamic magnetic field measurements, allowing for accurate characterization of field fluctuations contributing to pseudomotion [59, 121]. k‐space navigators, such as real‐time prospective shim correction, enable rapid field adjustments during acquisition, minimizing motion artifacts [122]. The DORK (Dynamic Off‐Resonance in k‐Space) method corrects respiration‐induced frequency shifts using phase information from navigator echoes, effectively mitigating image distortions like blurring and ghosting [116]. Additionally, Sensitivity Encoding (SENSE) Shimming exploits coil sensitivity variations to detect and compensate for B0 inhomogeneities, offering an efficient approach for dynamic field correction [123]. By integrating these techniques, fMRI acquisitions can more effectively disentangle true respiratory motion from field perturbation‐induced pseudomotion, thereby improving the accuracy and reliability of functional connectivity analyses. However, implementing these advanced correction strategies presents practical challenges, including increased computational demands and the need for specialized hardware, which may not always be feasible in standard fMRI protocols, thereby limiting their widespread application in large‐scale neuroimaging studies. Future work should focus on developing computationally efficient methods that address the aforementioned challenges while ensuring practical applicability in fMRI studies.

7.2. Cardiac and Motion Correction

Cardiac pulsations represent a high‐frequency physiological noise that poses a challenge for motion correction in fMRI due to aliasing at prevalent sampling rates. This aliasing arises as the cardiac frequencies frequently surpass the Nyquist frequency defined by typical fMRI temporal resolutions, leading to the potential for erroneous attributions to neural activity or head motion. To effectively disentangle these artifacts, there is a necessity for temporal resolutions that go beyond the conventional TRs, advocating for the utilization of faster imaging sequences that could provide sub‐second temporal resolution. A promising avenue for future research is the investigation of subjects across a spectrum of TRs, particularly employing ultra‐fast TRs [124]. Such a study design would facilitate the differentiation of the cardiac signal from the BOLD signal, providing a more unobstructed view of the underlying neural activity and potentially contributing to the refinement of motion correction algorithms.

8. Role of ML in Physiological Confound Correction

The integration of machine learning (ML) techniques with fMRI data analysis has significantly advanced the field of neuroimaging, especially in the context of physiological confound correction [26, 28, 29, 30, 31, 105, 118, 125, 126]. Among these advancements, automated ICA‐based correction methods have emerged as pivotal tools. The FIX method stands out as the most widely adopted solution within the ICA domain [26]. Rooted in ML algorithms, FIX adeptly distinguishes between ‘signal’ and ‘confound’ components derived from ICA‐decomposed fMRI data. Despite the impressive advancements brought about by automated ICA‐based correction methods like FIX, there are inherent limitations to consider. The reliance on manually‐selected features, while effective, introduces challenges in capturing the full spectrum of confounds, especially when dealing with data that deviate significantly from the training dataset. The specificity of features to the training data, such as those from the HCP, which is used to train the FIX, may limit FIX's generalizability across diverse datasets with varying acquisition parameters or subject populations. Moreover, the manual process of feature selection and the challenge of defining a universal criterion for confound‐related sources underscore the complexity and potential limitations of current automated ICA‐based correction methods.

To address the limitations inherent in manual feature engineering for physiological confound correction in fMRI data, Heo et al. [126] proposed an innovative deep learning framework. This framework leverages a dual convolutional neural network (CNN) approach to automatically extract spatial and temporal features from decomposed fMRI components. An attention mechanism within the model prioritizes critical features, mimicking expert human analysis without relying on predefined features. This method not only enhances accuracy and efficiency but also boasts broad applicability across diverse fMRI datasets, including those from infants and children, showcasing its potential to accommodate data variability better than traditional methods.

Yang et al. [125] developed a deep neural network (DNN) to enhance the quality of task‐based fMRI data by mitigating confounding factors without direct modeling. Diverging from traditional methods, this model does not depend on prior knowledge of the BOLD signal; instead, it employs a customized cost function. This function refines the model by maximizing the correlation difference between gray matter (task‐related) and white matter or cerebrospinal fluid (non‐task‐related) signals. When tested on both simulated and actual fMRI data, the DNN effectively reduces physiological noise and produces more uniform task‐response correlation maps in real data. Nevertheless, it is less suited for resting‐state fMRI data or when faced with head motion artifacts, owing to its dependency on task‐related designs and its challenge in addressing uniform confounding effects throughout the brain.

Physiological data is not routinely recorded in many fMRI experiments. Even in the fMRI studies where physiological data is collected, a significant proportion of the data is corrupted and unusable. As an example, up to 87% of the respiratory signals in the HCP‐D dataset are not usable [28]. Recent advancements have introduced ML algorithms to reconstruct these physiological measures directly from BOLD fMRI data, thus potentially obviating the need for direct physiological monitoring. In this context, Aslan et al. [127] developed a method employing a CNN to estimate cardiac waveforms directly from multislice fMRI datasets. The proposed model processes time‐series data, effectively denoising and reconstructing cardiac signals from raw fMRI‐derived inputs to produce high‐fidelity waveforms closely resembling plethysmogram recordings. This innovative approach not only enhances signal quality but also facilitates retrospective correction of physiological confounds without reliance on external physiological measurements.

Studies by Salas et al. [31], and Bayrak et al. [29, 30], utilized CNNs, U‐Net architectures, and Long Short‐Term Memory networks for the reconstruction of RV waveforms. These methods, validated using the HCP‐YA dataset, demonstrated promising results. Despite these advances, the initial models were limited in their ability to reconstruct the entire RV signal, particularly at the beginning and end of scans. To address these limitations, Addeh et al. [28] developed an approach employing three separate CNNs, facilitating the reconstruction of RV waveforms throughout the entire fMRI scans. The HCP‐D dataset proved instrumental in training and evaluating this model, given that pediatric populations are prone to increased motion artifacts and have reduced compliance with respiratory measurement protocols. In subsequent research, Addeh et al. [118] further refined this model by incorporating estimated head motion parameters alongside the BOLD signal within the CNN framework, substantially improving the fidelity of the reconstructed signals. Figure 8 presents examples of reconstructed RV waveforms over the entire scan duration, generated using the method proposed by Addeh et al. [118].

FIGURE 8.

Leveraging ML for enhanced physiological confound correction in fMRI. This figure compares measured and reconstructed RV waveforms over the entire scan duration. Panels (a)–(d) demonstrate the effectiveness of machine learning models, specifically CNNs, in reconstructing RV waveforms using BOLD signals alone (red line) and in combination with head motion parameters (blue line). The inclusion of head motion parameters significantly enhances the accuracy of the reconstructed signals, as evidenced by lower mean absolute error (MAE), mean squared error (MSE), and dynamic time warping (DTW) values and higher correlation values throughout the scan, especially in challenging regions marked by green and orange arrows. This example highlights the critical role of ML in improving the fidelity of physiological signal reconstruction, thereby aiding in the correction of physiological confounds in fMRI data. Adapted from Addeh et al. [118].

While ML methods like FIX represent a significant advancement in fMRI data analysis, there is a continuous effort toward developing more adaptable, generalizable, and automated methods for physiological confound correction.

9. Regional Sensitivity to Physiological Fluctuations

The magnitude of true neural‐related BOLD signal changes is typically minimal, often ranging from only 1% to 2% [21]. Physiological confounds, however, can exhibit magnitudes comparable to those of the true neural BOLD signals [18, 21]. Such confounds introduce significant artifacts that diminish the SNR, complicate the interpretation of data, and can potentially mislead statistical analyses at both individual and group levels during the investigation of neuronally‐related brain activation [26]. Dagli et al. [21] provided an example where physiological fluctuations, such as high‐frequency cardiac‐related confounds, if accounting for 25 to 30% of the total signal variance, can decrease the t‐statistic in tests comparing activated to baseline conditions by approximately 15%, highlighting the potential impact on statistical analyses in fMRI studies.

High‐frequency respiratory‐related confounds predominantly impact regions proximal to the brainstem and other medial areas, with notable respiratory noise observed in the superior sagittal sinus, and medial regions adjacent to the cingulum, precuneus, lingual gyrus, and areas near air‐tissue interfaces [18]. Conversely, high‐frequency cardiac fluctuations enhance BOLD signal components near major blood vessels, affecting the vertebrobasilar arterial system, the middle cerebral artery near the anterior temporal lobes and insula, and the anterior cerebral artery within the anterior interhemispheric fissure [18, 21, 85, 86]. These fluctuations influence approximately 27.5 ± 8.0% of voxels and are localized [21], driven by vessel‐dependent brain pulsatility. High‐frequency respiratory effects, in contrast, show a more extensive spatial distribution [18]. Low‐frequency cardiac confounds impact the occipital cortex, posterior cingulate, parietal lobes, and frontal regions. Notably, gray matter exhibits stronger overall MRI signal correlations with the heart rate timecourse than white matter, owing to its higher vascularity, resulting in greater BOLD signal variability and fluctuations [13].

P_ETCO₂‐related BOLD fluctuations are primarily concentrated in the gray matter, with significant changes observed in approximately 19% of gray matter and 14% of white matter [12]. These fluctuations are widespread but most prominently affect the occipital cortex, temporal cortex, posterior parietal cortex, insular cortex, frontal cortex as well as the sensory cortex [12, 24, 50, 51]. The larger fluctuations in the gray matter compared to the white matter likely reflect differences in metabolic activity, vascular regulation, and capillary density [12]. Other studies have also highlighted these effects: Rostrup et al. [128] found significant BOLD signal changes in the temporal and occipital gray matter in response to CO₂ challenges; Posse et al. [129] observed the greatest sensitivity to hypocapnia in the frontal, occipital, and parietal regions; and Kastrup et al. [43] reported larger BOLD sensitivity to breath‐hold in sensory‐motor and visual areas compared to the frontal areas, with even greater sensitivity noted in the cerebellum. The respiratory indices such as RV and RVT are believed to capture breathing‐induced changes in arterial CO₂, serving as a surrogate for P_ETCO₂. Although the brain regions that correlated with RV and RVT appear similar to those correlating with P_ETCO₂ [11, 14, 23, 50], the exact relationship between respiratory indices and CO₂ fluctuations, as well as how they jointly or distinctly explain components of the BOLD signal across common brain regions, remains unclear. Figure 9 illustrates the spatial distribution of brain regions influenced by physiological fluctuations, highlighting the distinct effects of low‐ and high‐frequency respiratory and cardiac variations.

FIGURE 9.

Regional Sensitivity to Physiological Fluctuations in fMRI. This figure illustrates the spatial distribution of physiological signal fluctuations affecting the BOLD signal across different brain regions, categorized by frequency and physiological origin. Low‐frequency respiratory‐related fluctuations (yellow) predominantly impact regions adjacent to large venous structures, including the superior sagittal sinus. Low‐frequency cardiac‐related fluctuations (blue) influence cortical and subcortical areas, notably the occipital cortex, posterior cingulate, and parietal lobes. High‐frequency respiratory‐related fluctuations (green) are primarily localized around the brainstem and medial regions such as the cingulum and precuneus, whereas high‐frequency cardiac‐related fluctuations (red) are concentrated near major arterial structures, including the vertebrobasilar system, the middle cerebral artery near the insula and anterior temporal lobes, and the anterior cerebral artery within the interhemispheric fissure. The anatomical slices (axial, coronal, and sagittal) provide a detailed visualization of these spatial patterns, highlighting the vessel‐dependent and region‐specific nature of physiological fluctuations. These maps were generated based on regions reported in the literature, rather than from original analyses. All images are displayed in the standard Montreal Neurological Institute (MNI 152) T1‐weighted 1‐mm brain template.

Variations in the regions affected by P_ETCO₂ compared to respiratory indices can be attributed to several factors highlighted in the literature. Firstly, prior studies have not consistently measured baseline values of hemodynamic parameters such as CBF, CBV, CMRO₂, OEF, and Hct. Discrepancies in the resting‐state values of these hemodynamic parameters can lead to divergent BOLD signal responses, potentially skewing interpretations of neural activity and physiological changes [1, 5, 10, 41]. Accurate measurement of these baselines using calibrated fMRI techniques would allow for effectively comparison of the effects of P_ETCO₂ and respiratory indices, thereby ensuring robust and reliable interpretation of fMRI data. Secondly, most studies have used a single hemodynamic response function with a fixed delay time between respiratory changes and the BOLD signal. For instance, Wise et al. [12] used a single delay of 6.3 s for all brain regions to analyze the relationship between P_ETCO₂ and the BOLD signal. However, regional variations in the delay of the BOLD response to respiratory processes have been reported, with delays to hypocapnia ranging from 5 to 12 s [129]. Accurate regional delay times are crucial for determining whether P_ETCO₂ or respiratory indices explain more signal variance, enhancing fMRI data interpretation and understanding of regional hemodynamic responses. Third, the impact of physiological noise is more pronounced at higher field strengths (e.g., 3 T vs. 1.5 T) because the BOLD signal strength—and consequently, the fraction of total noise—increases [27]. Studies employing different B₀ field strengths may therefore observe varying influences from physiological fluctuations. Fourth, thresholds for statistical significance, such as Z scores, vary across studies based on the specific goals and design considerations. For instance, Wise et al. [12] used a significance threshold of Z > 2.3, while Birn et al. [11], employed a stricter threshold of Z > 5.3. While these choices are typically justified by the study context, they nonetheless contribute to discrepancies in findings across the literature. Fifth, performance variability in P_ETCO₂ measurement and respiratory indices also plays a significant role. P_ETCO₂ can be noisier than respiratory indices due to both the sampling interval and the dependency on the subject's behavior [50]. On the other hand, respiratory indices (RV and RVT) may face issues such as peak detection errors, missing deep breaths, and inconsistent performance during breath‐holding [78], which can impact statistical analyses. The unrobust performance in different situations for both P_ETCO₂ and respiratory indices can lead to varying outcomes and create differences between brain regions affected by these measurements.

If respiratory indices primarily serve as predictors of CO₂ concentration, then a more direct measurement of CO₂ in the body, such as P_ETCO₂, may provide a more accurate prediction of physiological changes in the BOLD signal than respiratory indices. This is because RV and RVT reflect only chest expansion, approximating tidal volume while neglecting factors such as dead space and metabolic CO₂ output, making them indirect measurements of respiratory processes. Conversely, respiratory indices might explain additional variance in the BOLD signal beyond CO₂ levels alone. The expansion and contraction of the chest during breathing elicit a cascade of physiological processes, including modulation of CBF by intrathoracic pressure, heart rate, and corresponding autoregulatory mechanisms [130, 131]. Although CO₂ is a crucial intermediate variable, it may account for only a subset of these processes [50]. From a practical standpoint, if measurements derived from a respiration belt overlap significantly with those of P_ETCO₂ in explaining components of the BOLD signal, there may be limited additional benefit in monitoring P_ETCO₂ solely for reducing physiological noise. However, including both P_ETCO₂ and respiratory indices in the model generally explains more variance than either signal individually [24, 50], suggesting their complementary roles in physiological noise correction.

10. Summary

In this review, we explored the significant challenges posed by physiological confounds in fMRI studies and the diverse strategies developed to address them. Physiological variations, including respiratory and cardiac fluctuations, profoundly impact the BOLD signal, potentially leading to misleading interpretations of brain activity. Our review detailed the mechanisms underlying these confounds and their effects on fMRI data, highlighting the importance of accurate correction methods to enhance the reliability of fMRI results.

We examined the main categories of confound correction techniques: filtering methods, model‐based approaches, and data‐driven techniques. Filtering methods, while straightforward, often fail to accommodate the complex and dynamic nature of physiological signals. Model‐based approaches offer a more targeted method for addressing specific sources of confounds in fMRI studies without the risk of eliminating relevant signal data. However, this advantage comes with the requirement of simultaneous collection of physiological data during the fMRI scan, which may increase the complexity and cost of the study. Data‐driven methods, such as PCA and ICA, offer flexibility by identifying confounds directly from the data, although they require careful implementation to avoid removing genuine neural signals.

The integration of machine learning techniques has further advanced the field, enabling more automated and accurate correction of physiological confounds. Methods like FIX and deep learning frameworks have shown promising results in distinguishing between neural signals and physiological noise. However, these approaches also face challenges related to generalizability and the need for extensive training data.

Finally, we discussed the regional sensitivity of the BOLD signal to physiological fluctuations, emphasizing that different brain regions exhibit varying degrees of susceptibility to these confounds. This variability underscores the necessity for region‐specific correction strategies to ensure the accuracy of fMRI analyses. Moreover, different studies have employed various protocols and significance thresholds, complicating direct comparisons and making it challenging consolidate previous work. Establishing a standard protocol would facilitate more consistent and reliable results, allowing for better utilization and validation of past studies, but this is an open challenge with many aspects of fMRI.

In conclusion, addressing physiological confounds in fMRI is crucial for the accurate interpretation of brain activity. Continued advancements in correction methods, particularly those leveraging machine learning, hold the potential to significantly enhance the quality and reliability of fMRI studies without complicating data acquisition, paving the way for deeper insights into brain function and development.

Data Availability Statement

Data sharing is not applicable to this article as no new data were created or analyzed in this study.