3.1.1.

Treadmill walking

Fifteen studies investigating cortical activation during treadmill walking assessed baseline brain activation during quiet standing.^53–67 In contrast, Eggenberger et al.⁶⁸ chose slow walking ($2\mathrm{km}/\mathrm{h}$) for 1 min as the baseline condition. The duration of baseline cortical activation used for further analyses varied between 2.5^55,56 and 20 s^59–61 (for an overview see Table 1).

Table 1

Overview about the population, study designs, and data processing steps of reviewed fNIRS studies (note that the number of trials is reported per condition).

3.1.2.

Overground walking

Twenty-one studies conducting overground walking, quantified baseline brain activation in a standing position.^66,69–88 In contrast, two studies assessed baseline brain activation while walking⁸⁹ or during a predefined time period prior to a freezing of gait event (FOG; a sudden, brief inability to start movement or to continue rhythmic, repeated movements despite the internal intention to move).⁹⁰ The duration to assess baseline brain activity ranged between 5 s^69,73,76,81 and 5 min.⁷⁸ Most studies used 10 s to quantify baseline brain activity.^{66,70–72,74,75,82–88} Interestingly, Holtzer et al.^{74,75,83–85} asked their participants to conduct a simple counting task (in steps of 1) during the baseline condition (for an overview see Table 1).

3.1.3.

Postural tasks

In postural research, 13 studies assessed baseline brain activity during quiet standing.^91–103 The temporal duration to quantify baseline brain activity ranged from 2⁹³ to 60 s.⁹⁷ In most studies, data of 30 s^98–100 or a few seconds before starting the next trial^91,94,95,102 were used to assess baseline brain activation. In addition, Wang et al.¹⁰⁴ used 20 min quiet sitting to measure baseline connectivity (for an overview see Table 1).

3.2.1.

Treadmill walking

The studies that used a treadmill for the walking condition^{53–68,105–107} are shown in Table 1. Per task, a minimum of 2 trials¹⁰⁵ and a maximum of 10 trials^54–56 were performed. Most studies used three to five trials to assess task-relevant cortical activity.^{53,57,59–62,64,64,65,67,106,107} Task phases were set to 30 s in the majority of the studies,^{55–61,65,68,105–107} but Harada et al.,⁵³ Kim et al.,¹⁰⁶ and Mihara et al.⁵⁸ used 60 s, Koenraadt et al.⁵⁴ used 35 s, Metzger et al.⁶⁴ used 45 s, Suzuki et al.⁶⁷ used 40 s, Suzuki et al.⁶² used 90 s, and Fraser et al.⁶³ used 120 s. The time of the rest phases ranged in most studies between 25 and 60 s.^{53–57,59–62,65,68,106,107} Additionally, rest times of 15 s prior to^58,64 and after each walking trial⁵⁸ were reported while Suzuki et al.⁶⁷ implemented 10 to 25 s between trials (for an overview see Table 1).

3.2.2.

Overground walking

Twenty-three studies investigated cortical hemodynamic responses during overground walking.^66,69–90 For each condition, 1,^70,79 3,⁷⁶ 4,⁸⁷ 5,^{66,69,71,77,80,81,86} 6,^{73–75,82–85} and 15 walks were used.⁸⁹ Either the time for each task phase ranged between 10⁶⁹ and 120 s⁶⁶ or the participants were asked to walk a predetermined distance ranging between $\sim 4$^{73–75,82,83,85} and 90 m.⁸⁶ The resting phases prior to and after each trial lasted 20 s^77,79 or 60 s⁷⁶ and 10 s⁸⁷ or 30 s between the trials.⁷² Two studies used 20 s of rest between successive trials and 1 to 2 min of rest between successive task blocks.^80,81 Furthermore, in three studies, a rest of 2 min was used^66,86,90 while one study allowed participants to rest 5⁸⁸ or 30 min between tasks⁷¹ (for an overview see Table 1).

3.2.3.

Postural tasks

Regarding the examination of brain activity during a sensory organization test (SOT; a balance test using quantitatively different visual, proprioceptive, and vestibular cues to assess the quality of postural stance stability), two trials,⁹² three trials,⁹⁶ or four trials were conducted¹⁰³ which lasted 45,⁹² 40,¹⁰³ or 20 s.⁹⁶ The participants of the three studies using mechanical perturbations performed 15^95,102 to 30 trials⁹⁴ with a randomized perturbation duration of 5 to 20 s.^94,95,102 In semivirtual reality, seven trials with a task phase duration of 45 s were used.⁹⁸ The rest between task phases depended on the conducted tasks (see Table 1) and ranged between 4 and 20 s.^91,94,95,101 In other studies, a rest of 1¹⁰³ or 2 min was included.^92,98,99 To avoid fatigue, resting times after some trials that lasted a few minutes were common^91,92,94 (for an overview see Table 1).

5.1.1.

Baseline condition and duration

The majority of studies with walking or postural tasks assessed baseline brain activation in quiet standing. Interestingly, Holtzer et al.^74,75 used a silent counting task to avoid mind wandering. Mind wandering occurs up to 50% of the waking hours¹¹³ for instance during driving^114,115 especially when perceptual requirements are low.¹¹⁶ Moreover, the wandering of the mind is characterized by the processing of task unrelated thoughts such as worrying about the past or future,¹¹⁷ which evokes a stronger activation of default networks¹¹⁸ and hence changes the activation in PFC areas.^119,120 In addition, it was shown by Durantin et al.¹²⁰ that fNIRS is sensitive to detect mind wandering. Based on these assumptions, it is possible that mind wandering influences the cortical activation during baseline (and maybe motor control) affecting further analyzation processes. Hence, it might be advantageous to use the approach of Holtzer et al.,^74,75 which eventually minimizes the detrimental effect of mind wandering on cortical activation and leads to a more standardized baseline assessment. However, before the usage of this simple counting task can be recommended, further research should investigate its influence on cortical activation patterns including examination of enhanced reproducibility.

5.1.2.

Number and duration of trials and rest phases

Our results revealed that the number of trials and their durations varied across the studies evaluating walking or postural tasks. The most common time interval was set to 30 s. However, we are unaware of a study investigating the influence of measurement strategy (e.g., required number of trials to achieve a sufficient reproducibility). Hence, further methodological investigations to optimize fNIRS measurement protocols are needed. Moreover, the duration and number of the trials depend on the aim of the study. Longer measurement durations may be useful to study the contribution of different areas in the temporal course of movement execution. In contrast, longer measurement durations could result in motor fatigue. Motor fatigue does diminish performance for example in postural tasks^121–126 and would hence change underlying motor control processes. This again could potentially evoke altered hemodynamic responses, which were observed after cognitive fatigue.¹²⁷ However, research examining the interplay between a specific gross motor task and hemodynamic responses as a function of physical fatigue level has not been conducted yet.

Another interesting point influencing the trial duration is the combination of analysis methods. From a movement scientific view, the analysis of gait features (especially gait variability and stability) gives an insight in the central organization of motor control processes^128–131 and those are useful to detect risk groups such as fallers.^132,133 To reliably assess gait variability or stability, a larger number of strides is required^134,135 and as a consequence, a sufficiently long time period (in which an adequate number of strides can be undertaken) of the trial duration has to be recorded. The rest phase durations in included studies have varying temporal ranges. In general, empirical evidence suggests that refraction time or time with reduced responsiveness lasted for almost the same duration as stimulus time.¹³⁶ Hence, we recommend to include intertrial rest intervals with at least the same duration as the task period, especially in block design studies.

5.1.3.

Source–detector separation

The separation of source to detector is one important aspect for penetration depth^27,34 and the influence of extracerebral signals.^34,137 Our results indicated that 3 cm was the most commonly used distance in the reviewed studies. In the literature, different recommendations about optimal source–detector separation exist. While some authors recommend 4 cm,³⁴ other collectives recommend 3 cm.^138,139 In addition, especially in children or infants shorter interoptode distance ($>2.0\mathrm{cm}$) is recommended for usage.^22,139 The issue of the optimal separation between source and detector is a controversial debate because different third variables such as different colors of the participant’s skin and/or hair used wavelengths and head size could influence penetration depth.^34,140 Furthermore, the varying thickness of scalps, skulls, and cerebrospinal fluids in individuals and cortical regions^141–143 could influence the penetration depth and the sensitivity to hemodynamic changes in cortical layer.^142–144

Remarkably, a longer source–detector separation leads to a greater contribution of cerebral than extracerebral layer to obtain hemodynamics signals.^145–148 The penetration depth of light is less than half of the interoptode distance¹⁴⁷ causing short channel distances to cover only signals from noncerebral compartments.^137,141 For instance, at the source–detector separation of 3 cm, the contribution of the gray matter to the light absorption is estimated to range from about 20% to 30%.¹⁴⁹ Moreover, Kohri et al.¹⁵⁰ observed that at source–detector separation of 2, 3, and 4 cm, the cerebral tissue contributes to 33%, 55%, and 69% to the optical signal. Hence, we recommend that the source–detector separation should be greater than 3 cm to enhance the contribution of cerebral cortical layer to the optical signal.

5.1.4.

Placement of optodes

The majority of the studies used the 10 to 20 EEG systems to place the optodes. This standardized location system ensures the comparability among the different studies. The additionally used 3-D digitizer or individual MRI scan improves the registration of channels to specific brain areas. Based on the data we recommend for optode placement the usage of the 10 to 20 EEG systems to ensure the comparability among studies.

5.1.5.

Differential path length factor

Our results show that most studies used constant DPF with a value of 6. The usage of a constant DPF value seems not always appropriate because the brain undergoes age-related changes of gray and white matter,^151,152 intracranial volume,¹⁵³ and cerebral volume as well as blood flow¹⁵⁴, which may affect DPF.¹⁵⁵ Furthermore, methodological studies show that DPF values are (1) age-dependent and subject-specific,^110,155,156 (2) wavelength-dependent,^110,155,157 and (3) cortex region-dependent.^{110,155,158–160} Hence, it seems favorable to calculate specific DPF values to enhance the measurement accuracy in age-groups in which formulas to calculate age-specific DPF values are available (adults under 50 years).^110,155 Otherwise, “arbitrary units,”¹⁶¹ TOI,^162–164 or absolute values^137,163,165 could be used since those do not depend on a specific DPF value. In addition, it is suggested that the calculation of effect sizes is useful to deal with the DPF issue.¹⁶⁶ However, additional research is strongly needed that provides a formula to calculate DPF values for specific age-groups (adults older than 50 years) dependent on wavelength and cortex region. In our opinion, the optimal approach to quantify DPF, taking the dependency of DPF regarding subject, age, wavelength, and cortex region into account, is the direct quantification of DPF using frequency or time-domain NIRS.

5.1.6.

Data processing: signal filtering and movement artifact removal

In sum, either LPFs or HPFs were commonly applied in the reviewed studies to remove noise and drifts. Most of the studies used a cut-off frequency for LPF around 0.1 Hz and HPF around 0.01 Hz. The reviews of Brigadoi et al.,¹⁶⁷ Cooper et al.,¹⁶⁸ and Gervain et al.⁴⁰ recommended to use a bandpass filter (consisting of both LPFs and HPFs) with cut-off frequencies at 0.5 (LPF) and 0.01 Hz (HPF). The bandpass filtering should be used carefully to avoid accidental removal of stimulus-dependent hemodynamic response signals.¹¹¹ Hence, a higher cut-off frequency at 0.5 Hz (LPF) in conjunction with other more sophisticated filter methods is recommended to be used for the removal of movement and physiological noise.^111,167,168 Different methods such as PCA,^169–171 task-related component analysis,^172–174 CBSI,¹⁷⁵ wavelet-based filters,^{171,176–179} autoregressive algorithm-based filters,¹⁸⁰ Kalman filter,¹⁸¹ and Wiener filter¹⁸² are proposed for the filtering of fNIRS data. Interestingly, Nozawa et al.¹⁸³ suggested that effectiveness of motion correction filter methods depends on subject and task. However, reviews comparing a variety of filter methods recommend the additional application of wavelet filter^167,168 or spline technique.¹⁶⁸ These filter methods were occasionally applied in reviewed studies^80,105,106 leaving potential to optimize the filtering processes in further studies. Based on these assumptions, we recommend the usage of a bandpass filter and wavelet filter to reduce motion artifacts. If there are sudden shifts in the data (baseline shift), the approach developed by Scholkmann et al.¹⁸⁴ can be useful to remove them.

5.1.7.

Data processing: correction for physiological artifacts

Twelve studies recorded physiological signals such as heart rate, blood pressure, or arterial oxygenation saturation parallel to the fNIRS signals. Task-related systematic changes in heart rate, respiration rate, or blood pressure are known to influence the fNIRS signal and may cause false-positive results.⁴⁵ For instance, often unconsidered factors such as adding of speech as a task (e.g., in dual-task paradigms) lead to changes in partial pressure of end-tidal carbon dioxide, which influences cerebral hemodynamics and masked neuronal-induced activity changes.^185,186 Hence, to improve the accuracy of fNIRS, the recording and elimination of systemic physiological changes seems necessary.^45,187,188 The signals of additional physiological measures could be useful for filtering of fNIRS signal^189,190 or to ensure the absence of systematic physiological differences among the experimental conditions.⁸⁷ In addition, some measures such as heart rate variability could be used to study the interplay between the central (fNIRS) and the autonomic (e.g., heart rate variability) nervous system.^120,191 Furthermore, filter methods based on PCA and independent component analysis, which were applied in six studies,^{55,56,73,76,100,102} could be used to remove movement-related¹⁶⁷ or physiological artifacts.^{169,170,192–195} In addition to the other filter methods,^196,197 a more “direct” approach to reduce extracerebral noise is the use of short separation channels or multidistance technique^198–200, which were applied in only two of the reviewed studies.^54,68 Short separation channels have a small distance between source and detector to record extra cerebral signals, such as superficial blood flow.^141,198,201 These extracerebral signals are used to filter the remaining fNIRS data. Previous studies revealed that the application of short separation channels is powerful in reducing extracerebral noise^{141,145,200,202–208}, which contaminates fNIRS signals.^{45,199,201,209–214} The optimal distance between short separation channels varied across different cortex regions^141,202 but should be generally $<1\mathrm{cm}$ for measurement on the head of adult humans. Hence, further development and implementation of short separation channels (multidistance technique) could enhance the accuracy of fNIRS measurements and have to be considered whenever technically possible.

5.1.8.

Data processing: final data processing and statistical analysis

Most studies used baseline normalization and baseline correction to circumvent the influence of different path lengths factors.¹⁶⁶ Furthermore, averaging of channels across trials and in specified ROIs was common practice in the reviewed studies.

Some studies divided their task phase in different time periods, which seems useful for studying the contribution of cortical areas in different temporal periods during task execution. Therefore, attention should be paid to the temporal delay of $\sim 2$ to 5 s in hemodynamic response.^69,107,139

The majority of the reviewed studies used simple statistics based on processing mean values over the task period. This approach, however, tends to result in a loss of acquisition of information because it does not consider the temporal shape of the fNIRS signal.¹⁹² Hence, some authors suggest that the analysis of fNIRS data with general linear models is more favorable.^192,215 However, the choice of the statistical analysis methods should depend on the research question and the experimental design.²¹⁶ For instance, in an event-related design, the application of a general linear model is a valid technique²¹⁶ whereas simple statistics might also be appropriate (and commonly used¹⁹²) especially in studies utilizing block designs.^{55,56,59–62,107,217} The majority of reviewed studies used parametric methods for statistical data analysis. In fNIRS studies, the assumptions for parametric tests are sometimes violated (e.g., normal distribution due to small sample size); therefore, nonparametric tests are a considerable option.^218,219 Moreover, nonparametric tests are more robust and less influenced by outliers or nonnormal distributed data^220–222 and are recommended to use in fNIRS studies. From another point of view, in neuroscience, multiple experimental conditions (crossed) or multiple observations per condition (nested) were used.^223,224 Furthermore, different categorical or continuous confounding variables have to be considered (e.g., gait speed, education, and gender) and/or data were unbalanced or incomplete, which makes it necessary to use advanced statistical methods.^223,225 To account for those problems, linear mixed-effect models can be used.^10,224–226 However, statistical methods should be chosen carefully considering the experimental design and distribution of recorded data. A further description of statistical methods for fNIRS data is given in the reviews of Tak and Ye¹⁹² and Kamran et al.²²⁷

5.1.9.

Markers for the assessment of cortical activation

The majority of reviewed studies used only oxyHb for the quantification of cortical activation since a change in oxyHb is assumed to be a more robust marker of changes in regional cerebral blood flow than changes in deoxyHb.^160,228,229 However, this procedure seems questionable because neuronal activity is not just mirrored in an increase of oxyHb but also in a decrease in deoxyHb in healthy adults.^30,230 Furthermore, an enhanced level of physiological noise is more prominent in oxyHb signals³⁰ and the decrease in deoxyHb is related to an increase in BOLD contrast obtained in fMRI^231,232, which supports the validity of the evaluation of deoxyHb changes. In pathological states, neurovascular coupling might perhaps be impaired, which results in altered concentration changes in deoxyHb during neural activity.²³⁰ Lindauer et al.²³⁰ assumed that in some pathological states, an increase in deoxyHb may reflect neural activity. Based on the mentioned assumptions, it seems favorable to report at least oxyHb as well as deoxyHb to assess task-dependent activity.^30,45,165

5.2.1.

Walking

Evidence from neuroimaging studies point out that two distinct supraspinal locomotor networks are responsible for the control of walking and standing^1,233–237 (see Fig. 3). The direct locomotor network consists of the primary motor cortex (M 1) and the cerebellar locomotor region and is potentially activated in the absence of pathologies or challenging situations.²³⁵ In the indirect locomotor pathway, the neuronal commands are transmitted via PFC and SMA to the basal ganglia and subthalamic as well as mesencephalic locomotor regions.^233–237 The indirect locomotor pathway becomes activated when the automatic execution of walking is impaired (e.g., in challenging situations) and compensatory mechanisms are necessary.^44,238 This assumption is supported by findings of our reviewed fNIRS studies, which reported more pronounced activation in prefrontal structures in (1) in adults during dual-task walking,^{57,63,63,66,69,72–74,77,86,89} (2) in adults during fast walking,^53,64 (3) in obese persons,⁸⁷ (4) in individuals with low gait capacity during fast walking,⁵³ (5) in older adults with high level of perceived fatigue⁸⁵ or stress,⁸⁴ (6) in old adults with increased fall risk,⁸³ and (7) in neurological patients.^{58,70,71,75,80,82,90} Remarkably, the PFC activation in neurological patients correlates with their step widths,⁷¹ which again (1) is associated with balance control²³⁹ and (2) serves as a predictor of falls.²⁴⁰ Furthermore, correlations between cortical activation and motor performance,^55,56 especially obvious in dual-task walking conditions,^76,77,89,105 was observed. This reinforces the important role of cortical areas in motor control. Moreover, the reduction of PFC activity after a motor-cognitive intervention program (lasting 8 weeks)⁶⁸ perhaps originated from the shift toward a more automatic control of locomotion relying on the enhanced usage of direct locomotor pathway via M1, cerebellum, and spinal cord.^{1,233,234,238}

Fig. 3

Schematic illustration of the indirect and direct locomotor pathways as a function of the degree of automaticity in motor control.

However, premotor areas and the SMA play a role in different cognitive processes^241–243 and were activated as a function of task difficulty in a variety of cognitive domains.^244–246 Hence, the phenomenon of a more pronounced activation of premotor areas (as part of indirect locomotor pathway) in diseased cohorts (or during challenging motor tasks) is perhaps not fully attributable to motor task complexity but partly also to general task complexity.

However, the decrease in PFC activity in a complex visual task¹⁰⁵ or difficult working memory tasks during walking^79,88 may not be induced by the shifts in locomotor pathways but rather originate from the prioritization of task-relevant areas as consequence of the limited resources of the brain.²⁴⁷ While those three studies focused only on PFC activity, it is difficult to draw a final conclusion about potentially underlying cortical processes in other areas. Hence, to elucidate the mechanisms with respect to task prioritizations, we require further research^248,249 including the simultaneous assessment of more cortical structures (e.g., motor areas).

For the design and monitoring of rehabilitative interventions, fNIRS could be a promising tool.⁴² For instance, the SMC activity decreases during weight-supported walking in stroke patients⁵⁹ and could be a hint that weight supports lower task complexity.²⁵⁰ Interestingly, a verbal preadvice^67,94 or the usage of mechanical assistance during walking^61,106 increases central nervous load. These findings could be useful to create tailored rehabilitation programs that consider mental load as variable for workload assessment.

5.2.2.

Postural tasks

As pointed out for walking, neural control of posture is realized via direct or indirect pathway²⁵¹ which are shown in Fig. 3. Our results reveal that the PFC activation is enhanced in (1) neurological patients during standing⁹³ or during postural perturbations^95,102 and (2) healthy adults during challenging balance tasks.^91,98,99 These findings and the observations that PFC activity and SMA are associated with balance measures^95,100,102 support the notion that indirect locomotor pathway is crucial for neuromotor processes in nonautomatized challenging situations.

Additionally, altered sensory information evoked by the execution of SOT induces a higher activation especially in STG.^92,96 The STG is associated with (1) the control of more difficult balance tasks,⁹⁷ (2) the integration of vestibular information,^252–254 and (3) the spatial orientation.²⁵⁵ So far, the mentioned studies did include only young participants.^92,96 While aging changes the contribution of somatosensory, vestibular, and visual system in balance tasks,²⁵⁶ it seems necessary to enlarge existing knowledge about cortical sensory integration processes.