Shedding light on pain for the clinic: a comprehensive review of using functional near-infrared spectroscopy to monitor its process in the brain

Introduction

The pain experience profoundly impacts patients’ quality of life [59]. The estimated economic impact of pain, from direct medical costs to loss of productive time, is $600 billion every year [41]. Nevertheless, in clinics, patients are faced with the highly subjective task of reporting their pain on an arbitrary 0 to 10 pain scale, where 0 represents ‘no pain,’ and 10 represents ‘Pain as bad as you can imagine’ [127]. Though collectively evaluated with other multi-dimensional questionnaires like McGill and even affective ones as PANAS [86,165], the visual analogue scale (VAS) lacks specificity and often correlates with subjects’ positive and negative emotions at the time of rating [97].

An array of functional neuroimaging techniques has been developed in the past decades to not only understand the neuronal signature of pain, but to potentially visualize its ongoing process in the brain. These functional neuroimaging techniques, including functional magnetic resonance imaging (fMRI) [22,26,81,152,153,162], positron emission tomography (PET) [26,32,61,82], functional near-infrared spectroscopy (fNIRS) [17,37,42], electrical-magnetic property-based techniques electroencephalography (EEG) and magnetoencephalography (MEG) [26,62,112,117,138,140], demonstrated the feasibility of studying pain perception in the brain in vivo. Among these imaging modalities, fNIRS is an emerging optical technique that detects regional cerebral blood flow by gauging near-infrared light absorption variation, which has a wavelength between 650 −1000 nm [39,139]. fNIRS has a higher temporal resolution (~ 200 Hz) than fMRI, and higher spatial resolution (~ 1 cm) than EEG [84,136,139]. fNIRS has a few additional crucial advantages over other neuroimaging techniques that facilitate its application for the clinical environment, even for the pediatric population that is notoriously difficult for neuroimaging data collection. First, it is portable, safe, and relatively motion tolerant. Besides, it operates quietly and does not require ionizing radiation or drug injection (e.g., radiotracers), compared with other functional neuroimaging modalities [11,39,139]. Finally, the fNIRS signal is not vulnerable to electrical and magnetic field changes [136].

fNIRS was first applied to the newborn population in the clinical environment in 1995 [23] due to the challenges of studying the younger population with other imaging modalities like MRI and EEG [14,123,134]. In 2008, researchers investigated thermal heat stimulation evoked acute pain in healthy subjects with a large matrix of fNIRS probes monitoring bilateral prefrontal and somatosensory cortices [13]. In 2015, for the first time, machine learning/artificial intelligence algorithms were applied to fNIRS data collected from healthy subjects undergoing hand immersion in cold water. More recently, our group merged clinical augmented reality (AR) and artificial intelligence (AI) with fNIRS imaging, a concept called CLARAi [54]. The study demonstrated a preliminary framework that can predict and visualize patients’ cortical activation patterns associated with pain suffering and its location in real-time based on random sampled fNIRS data that were collected from patients with hypersensitive teeth.

This trend of studies demonstrated the potential of near-infrared light in pain research and objective assessment. Although fMRI, PET, EEG, and MEG are more common non-invasive imaging modalities for pain research, the number of fNIRS-based pain studies keeps increasing. Such increase is partially due to the portability and insensitivity to the magnetic and electric field that granted fNIRS the ability of real-time pain surveillance in many complex scenarios, including the patient’s bedside and pediatric applications. This article aims to consolidate the study pipeline for using fNIRS to image pain in the brain by providing a comprehensive review of 69 research projects published between 1995 and 2019. We specifically focused on the fNIRS relevant experimental design, probe localization, data collection, data processing, and primary findings (summarized in Table 1). Finally, we will discuss the future trend of using fNIRS combined with artificial intelligence and augmented reality for pain studies in the clinical environment.

Table 1.

Pain neuroimaging studies using fNRIS.

Study	Objective (study cortical hemodynamic response)	Study design	Brain region detection (S-D distance)	Population(Control sample size)	Outcome (Primary)	Limitations/Missing data	NOS/ RoB2 Score	Review disagreement
Bucher et al., 1995 [23]	Heel lance after sucrose administration	RCT	Temporal cortex (4 cm)	N=16 (Infant)	Decreased [HbO] and [HbR] after sucrose administration compared with the control group.	Insufficient data to determine the significance	6 (RoB2)
Kussman et al., 2005 [73]	biventricular repair surgery	NRS	PFC (3/4 cm)	N=62 (Infant)	rSO2 increased during surgery.	Insufficient data to determine the significance	6 (NOS)
Akın & Bilensoy, 2006 [4]	Breath-holding in migraineurs	NRS	PFC (2.5 cm)	N=6 (NC=6)	Suppressed [HbO] and [HbR] responses in migraine patients compared to healthy controls.	Small sample size	7 (NOS)
Akin et al., 2006 [5]	Breath-holding in migraineurs	NRS	PFC (2.5 cm)	N=6 (NC=6)	The initial dip and recovery of [HbO], [HbR], and [HbT] responses in migraine patients were ten times lower than healthy control.	Small sample size	7 (NOS)
Bartocci et al., 2006 [10]	Venipuncture	NRS	S1 (4 cm)	N=29 (NC=11) (Infant)	[HbO] increase in the S1. Male responses were greater.	Limited fNIRS measurement coverage	7 (NOS)
Slater et al., 2006 [142]	Heel lance procedure	NRS	S1	N=18 (NC=11) (Infant)	[HbT] increase in the S1.	Limited fNIRS measurement coverage	8 (NOS)
Becerra et al., 2008 [13]	Heat stimulation to the hand	NRS	S1/PFC (3cm)	N=9	[HbO] increase in bilateral PFC and S1.	Small sample size	7 (NOS)
Gelinas et al., 2010 [42]	Surgical pain during cardiac surgery	NRS	PFC (3/4 cm)	N=40	rSO2 increase in the PFC during surgery.	A low range of Variability in results	8 (NOS)
Muller et al., 2010 [95]	Breast cancer surgery	NRS	PFC	N=24	rSO2 increase in the PFC after surgery.	Low data sampling rate	6 (NOS)
Viola et al., 2010 [158]	Migraine with aura	NRS	PFC/S1/Parietal/Occipital	N=8	rSO2 increase in all regions after a migraine attack.	Lack of method details	6 (NOS)
Ozawa et al., 2011 [106]	Hand laterality of venipuncture	RCT	PFC (3 cm)	N=21 (NC=19) (Infant)	[HbO] increase at PFC. Right-hand venipuncture evoked a higher response.	Limited fNIRS measurement coverage	6 (RoB2)	Randomization process
Ozawa et al., 2011 [107]	Repeated venipuncture	NRS	PFC (3 cm)	N=80	Prefrontal cortical activity in response to pain correlated with emotional or stress responses.	Limited fNIRS measurement coverage	7 (NOS)
Watanabe et al., 2011 [164]	Sumatriptan injection during a migraine attack	NRS	S1/M1	N=4 (NC=4)	Decreased [HbO] after sumatriptan injection in migraine patients.	Small sample size	8(NOS)
Barati et al., 2013 [7]	Cold pressor test	NRS	aPFC (2.8 cm)	N = 20	[HbT] increase in the aPFC, but right more than left.	Limited fNIRS measurement coverage	6 (NOS)
Lee et al., 2013 [76]	Mechanical force applied to a finger	NRS	PFC/S1 (3 cm)	N=7	[HbO] increase in the PFC and S1.	Small sample size	6 (NOS)
Ranger et al., 2013 [122]	Chest drain removal following cardiac surgery	NRS	S1 (4 cm)	N=20	[HBR] increase in the right-side S1.	Limited fNIRS measurement coverage	6 (NOS)
Holper et al., 2014 [49]	Mechanical pressure at the lower back	NRS	aPFC (2.0/2.5/3.5/4.0 cm)	N=13	[HbO], [HbT]decrease, [HbR] increase at aPFC by left paravertebral muscle stimulation. [HbT] increase at aPFC by right paravertebral muscle stimulation.	Limited fNIRS measurement coverage	6 (NOS)
Sakuma et al., 2014 [131]	Mechanical pain: periodontal pocket probe on the gingiva	NRS	PFC (3 cm)	N=23	[HbO] decrease in the left aPFC and DLPFC.	No female participant	6 (NOS)
Hwang and Seol, 2015 [58]	Automatic and manual heel lance	RCT	aPFC	N=12 (infant)	rSO2 decrease for both groups, but less in the automatic heel lance process.	Lack of fNIRS set-up details	7 (RoB2)
Muthalib et al., 2015 [96]	Electrical muscle stimulation	NRS	PFC/SMA/S1 (3 cm)	N=9	[HbO] increase and [HbR] decrease at SMA, S1, S2, M1.	Small sample size No female participant	6 (NOS)
Ono et al., 2015 [104]	Dental occlusal discomfort	NRS	PFC	N=25	[HbO] increase in the right DLPFC for severe dental occlusal discomfort.		6 (NOS)
Pourshoghi et al., 2015 [118]	Migraine patients receiving an infusion of magnesium sulfate, valproate sodium, and dihydroergotamine	NRS	aPFC	N=41	[HbT] increase after magnesium sulfate infusion, while decrease after dihydroergotamine.	Limited fNIRS measurement coverage	7 (NOS)
Racek et al., 2015 [120]	Hypersensitive dental pain	NRS	PFC/S1 (2.7 cm)	N=21	[HbO] increase at bilateral PFC and S1.	13 % of the trials were excluded	6 (NOS)
Rojas et al., 2015 [129]	Acupuncture needle operation	NRS	S1/M1	N=6	[HbO] and [HbT] increase at S1.	Small sample size	6 (NOS)
Uceyler et al., 2015 [157]	Pressure stimulation at the dorsal forearm (FM)	NRS	PFC/S1/M1	N=25 (NC1=10) (NC2=35)	Stronger increases in [HbO] in the bilateral DLPFC for the FM group to pain stimulus in contrast to healthy controls.		8 (NOS)
Yucel et al., 2015 [175]	Electrical stimulation at 5 Hz	NRS	aPFC/S1/M1 (3cm)	N=11	[HbO] increase and [HbR] decrease at bilateral S1. [HbO] decrease at aPFC.	Small sample size	6 (NOS)
Aasted et al., 2016 [2]	Electrical stimulation at 5 Hz	NRS	PFC (3 cm)	N=10	[HbO] decrease at PFC.	Small sample size	6 (NOS)
Becerra et al., 2016 [12]	Colonoscopy	NRS	aPFC/S1/M1	N=17	[HbO] decrease in the (left) anterior prefrontal cortex to the insufflation response.	23.5 % of the data are excluded	6 (NOS)
Bembich et al., 2016 [15]	Heel lance procedure	NRS	posterior frontal cortex /S1/M1 (2cm)	N=16	[HbO] increase in the posterior frontal cortex, contralateral to the side of the heel lance procedure.		6 (NOS)	Assessment of outcome
Bembich et al., 2016 [16]	Mother with postpartum depression watching their own newborn’s pain	NRS	S1/M1	N=30	The empathy of pain evoked [HbO] increase in the left S1 and right superior temporal cortex.		7 (NOS)
Kussman et al., 2016 [72]	Catheter ablation of arrhythmias	NRS	S1/aPFC (3 cm)	N=8	[HbO] decrease in the aPFC.	Small sample size	6 (NOS)
Pourshoghi et al., 2016 [119]	Cold pressor test	NRS	aPFC (2.8 cm)	N=19	[HbO] increase in the aPFC.	Limited fNIRS measurement coverage	6 (NOS)
Vrana et al., 2016 [159]	Mechanical pressure at the lower back (chronic lower back pain)	NRS	S1/SMA (2.5cm - 4.5cm)	N=14 (NC=22)	[HbO] increase in the right S1 for patients with chronic lower back pain.	9% HC data excluded due to low SNR	8 (NOS)
Vrana et al., 2016 [160]	Mechanical pressure at the lower back	NRS	S1/SMA (2.5 cm – 4.5 cm)	N=22	[HbO] increase in the right S1 and SMA.	9% HC data excluded due to low SNR	6 (NOS)
Yennu et al., 2016 [174]	Thermal heat stimulation respectively at right/left forearm and right TMJ	NRS	PFC (3 cm)	N=16 (NC1=9) (NC2=9)	[HbO] increase in the left PFC for right forearm stimulation. [HbO] increase in the bilateral PFC for right TMJ and left forearm stimulation.		7 9(NOS)
Barati et al., 2017 [8]	Cold pressor test	NRS	aPFC (2.8cm)	N=21	[HbO] increase and [HbR] decrease in the aPFC.	Limited fNIRS measurement coverage	6 (NOS)
Eken et al., 2017 [35]	Median nerve and thumb stimulation with TENS (FM)	NRS	S1/M1	N=19 (N=17)	FM patients showed higher [HbO] than HC after median nerve, left and right-hand stimulation.		7 (NOS)
Hong et al., 2017 [51]	Hand poking and cold temperature stimulation	NRS	S1/M1 (2.8 cm)	N=8	[HbO] increase in the S1.	Small sample size	6 (NOS)
Meyer-Frießem et al., 2017 [87]	Electrical stimulation on the right ventral arm	NRS	aPFC	N=20	No significant rSO difference between painful and non-painful stimulations.	No female participant	6 (NOS)
Morikawa et al., 2017 [92]	Ischemic compression applied to myofascial trigger points on chronic neck pain	RCT	aPFC (3 cm)	N=11 (NC=10)	Compression at myofascial trigger points decreased [HbO] in the DMPFC.	No male patient	6 (RoB2)
Mukaihara et al., 2017 [94]	Combination of a paravertebral block and general anesthesia during thoracotomy	RCT	aPFC (3.7/4.3 cm)	N=17 (NC=17)	[HbO] increase in the general anesthesia group, but not in the combined paravertebral block and general anesthesia group.		6 (RoB2)
Rojas et al., 2017 [37]	Thermal stimulation on left hand	NRS	S1/M1 (3 cm)	N=18	[HbO] increased at S1.		6 (NOS)
Ren et al., 2017 [126]	Verbal fluency task (somatoform pain disorder)	NRS	PFC (3cm)	N=24 (NC=24)	Depressed [HbO] increases were found in SPD patients during VFT compared with healthy controls.		8 (NOS)
Suemori et al., 2017 [146]	Pediatric cardiac surgery	NRS	aPFC	N=399	[HbO] decrease and [HbR] increase	Not a single type of task (surgery) included	6 (NOS)
Chou et al., 2018 [29]	Verbal fluency task (FM)	NRS	PFC/S1	N=11 (NC=13)	Depressed [HbO] increases were found in FM patients during VFT compared with healthy controls.		7 (NOS)
Eken et al., 2018 [34]	Finger tapping task and median nerve stimulation (FM)	NRS	S1/M1	N=19 (NC=17)	Higher [HbO] during FTT, and lower [HbO] during MNS, in healthy control than FM patients.		7 (NOS)
Eken et al., 2018 [36]	Median nerve and thumb stimulation with TENS (FM)	NRS	S1/M1	N=19 (NC=17)	FM patients showed higher [HbO] than HC after median nerve, left and right-hand stimulation.		8 (NOS)
Frie et al., 2018 [40]	Chemical smell stimulation	NRS	PFC/S1/OC (2–4 cm)	N=17 NC1=15 NC2=12 (Infants)	[HbO] increase at S1 in full-term and very preterm infants.		8 (NOS)
Hu et al., 2018 [52]	Hypersensitive dental pain	NRS	PFC/S1 (2.7 cm)	N=21	[HbO] and RSFC modulated subjective pain experience.	43% data excluded from the analysis	6 (NOS)
Hucke et al., 2018 [56]	Irritating chemical stimulations of the nasal divisions of the trigeminal nerve	NRS	S1/M1	N=21 (NC=14)	[HbO] increase in the S1.	26% data excluded from the analysis	8 (NOS)
Olbrecht et al., 2018 [101]	Electrical current stimulation	NRS	PFC (2.5 cm)	N=91 (Children)	[HbO], [HbR], and [HbT] increase.	18% data excluded from the analysis	8 (NOS)
Peng et al., 2018 [109]	Electrical stimulation	NRS	aPFC/S1 (3 cm)	N=16	[HbO] decrease at aPFC.	12.5% data excluded from the analysis	6 (NOS)
Peng et al., 2018 [109]	Electrical stimulation at 5 Hz under immediate-release morphine condition	RCT	PFC/Right S1 (3 cm)	N=14	[HbO] decrease at medial aPFC. [HbO] increase at contralateral S1. Morphine attenuated the magnitudes of [HbO].	21% data excluded from the analysis	5 (RoB2)	Randomization process
Rioualen et al., 2018 [128]	Venipuncture with breastfeeding and sucrose administration	RCT	S1	N=114 (Infants)	[HbT] increase in the S1. No difference between breastfeeding and sucrose administration.	1% data excluded from the analysis	6 (RoB2)	Randomization process
Roue et al., 2018 [130]	Venipuncture	NRS	S1	N=114 (Infants)	[HbO], [HbT] increase in the S1.	11% data excluded from the analysis	6 (NOS)
Wolff et al., 2018 [169]	Static muscular endurance task with different self-regulation strategies	RCT	PFC (3 cm)	N=30 (NC=30)	[HbO] increase and [HbR] decrease in the DLPFC and VLPFC		7 (RoB2)
Wriessneggr et al., 2018 [170]	Imagery pain condition	NRS	PFC/SMA/S1 (3 cm)	N=20	[HbO] increase at left S1.		6 (NOS)
Xie et al., 2018 [171]	Empathy for pain task	NRS	PFC/Left S1	N=16 (NC=16)	[HbO] increase at left DLPFC, ACC, left S1, and M1.		7 (NOS)
Chen et al., 2019 [28]	Breath holding task (FM)	NRS	PFC (3/3.5/4 cm)	N=27 (NC=26)	Less [HbO] increase and [HbR] decrease during the breath holding task in FM patients.		7 (NOS)
Gokcay et al., 2019 [45]	Figure tapping and TENS (FM)	NRS	S1/M1	N=19 (NC=16)	Classification sensitivity ranges between 0.64–0.85, specificity ranges between 0.73–0.92.		6 (NOS)
Hu et al., 2019 [54]	Hypersensitive dental pain	NRS	S1/PFC (2.7 cm)	N=21	[HbO] decrease in the PFC and increase in the S1.	43% data excluded from analysis	6 (NOS)
Kodama et al., 2019 [68]	Compression of myofascial trigger points (chronic low back pain)	RCT	PFC/SMA/S1 (3 cm)	N=16 (NC=16)	[HbO] and [HbT] decrease in the PFC for the compression of myofascial trigger points. While [HbO] and [HbT] increase in the compression of non-myofascial trigger points		6 (RoB2)
Rojas et al., 2019 [38]	Acupuncture- needle operation	NRS	S1/M1	N=11	[HbO] increased during needle insertion, twirl and removal.		6 (NOS)
Sawosz et al. 2019 [135]	Surgery/Laparoscopy	NRS	aPFC (3 cm)	N=17	[HbO] and [HbR] increase.	No male participants. 12% data excluded from analysis	6 (NOS)
Schaal et al. 2019 [137]	Cold pressor test and stressful mental arithmetic task	NRS	PFC (3 cm)	N=40	Stressful mental arithmetic task caused increased [HbO] in the PFC bilaterally, cold pressor test caused decreased [HbO] in the left PFC.		6 (NOS)
Sharini et al., 2019 [141]	Cold pressor test	NRS	PFC (3 cm)	N=12	[HbO] and [HbR] increase		6 (NOS)
Sun et al., 2019 [147]	TMS on neuropathic pain after spinal cord injury	RCT	aPFC/S1/M1	N=14 (NC=7)	TMS treatment for at least two weeks suppressed handgrip task-induced [HbO] in the PMC and M1.	19% participants excluded	6 (RoB2)	Randomization process/ Intervention
Wang et al., 2019 [163]	Inter-personal cooperation and competition under acute pain condition (capsaicin on dominant forearm)	RCT	PFC/right S1 and M1	N=32 (NC=34)	Bilateral prefrontal and right parietal [HbO] synchronization between pairs of participants.	No male participants	8 (RoB2)
Gentile et al., 2019 [43]	Pre-stimulation resting state and slow/fast finger tapping task (FM)	NRS	S1/M1 (3cm)	N=24 (NC=24)	Similar RSFC, Lower [HbO] increases in the FM group than the HC group.		7 (NOS)

S1: somatosensory cortex, S2: secondary sensory cortex, M1: motor cortex, PFC: prefrontal cortex, aPFC: anterior prefrontal cortex, DLPFC: dorsal lateral prefrontal cortex, VLPFC: ventrolateral prefrontal cortex, DMPFC: dorsomedial prefrontal cortex. SMA: supplementary motor cortex, PMC: premotor cortex, OC: primary olfactory cortex. [HbO]: oxygenated hemoglobin concentration change, [HbR]: de-oxygenated hemoglobin concentration change, [HbT]: total hemoglobin (oxygenated + de-oxygenated hemoglobin) concentration change, RSFC: resting-state functional connectivity, S-D: source-detector, TMJ: temporomandibular joint, FM: fibromyalgia, HC: healthy control, rSO2: regional cerebral oxygen saturation, TENS: transcranial electrical nerve stimulation, RSFC: resting state functional connectivity, NRS: Non-randomized controlled study, RCT: randomized controlled trials, NOS: Newcastle-Ottawa Scale, CRBT: Cochrane Risk of Bias Tool, SNR: signal-to-noise ratio, N: number of participants, NC: number of control participants. S-D distance: source-detector distance.

Literature Search Method and Inclusion Criteria

We conducted the searches at the end of 2019 and queried three scientific databases, the Web of Science, PubMed, and Scopus. The following terms or combinations were used for the article search: [ near infrared spectroscopy | fNIRS | NIRS] & [ pain | nociception | chronic pain ] & brain. The language restriction was set to “articles written in only English.” As a backup, we also searched the Google Scholar database to prevent literature omission at our best. We then reviewed each article following the predefined inclusion criteria: 1. Published in a peer-reviewed journal and reviewed by at least two review authors in the current study, 2. Used fNIRS as the neuroimaging technique, 3. Empirically studied a pain condition in humans, 4. The research sample size (case + control) included must be greater than 5 (N>5), excluding case-study.

Study Quality Evaluation

To evaluate the risk of bias for the studies collected in this review, we employed different evaluation methods based on the study design type. We used the Revised Cochrane risk-of-bias tool for randomized trials (RoB2) [144] to evaluate the quality of randomized control trials (RCT), including the cross-over design and the parallel design. The Newcastle-Ottawa quality assessment scale (NOS) [166] was used to evaluate the quality of non-randomized studies (NRS), including the non-randomized control study, prospective cohort study, and the case-control study. The RoB2 scale (5–15) is the lower the better, while the NOS scale (0–9) is the higher the better. We also predefined thresholds to exclude studies with less optimal qualities. A study was excluded if the RoB2 score is higher than 10 for RCT studies, or the NOS score is lower than 6 for the NRS studies. At least two review authors completed the evaluation of each study. The disagreements encountered by the review authors were resolved by consensus. We also reported the disagreements associated with each article (if any) in Table 1. We manually went over each paper and excluded studies that did not report detailed fNIRS set-up, data collection, and analysis. Ultimately, we collected 69 articles to be illustrated in the current review article. A flowchart following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines that describe the article selection process is presented in Figure 1.

Figure 1.

Article search and selection process.

fNIRS experiment protocol design

Pain studies usually have a small sample size due to the difficulty in participant recruitment and the high cost of neuroimaging technology. Therefore, experiments need to be carefully designed to maximize the contrast between pain and the baseline. Besides, researchers need to be cautious about the specificity of pain evoked brain activation patterns, meaning a pain-related neuro-signature must be detected when there is pain reported and must not be detected when there is no pain.

There are two types of pain – acute and chronic pain, which call for different experimental designs. Acute pain is usually defined as “the normal, predicted physiological response to an adverse chemical, thermal or mechanical stimulus, associated with surgery, trauma, and acute illness.” [25]. Chronic pain is defined as a pain condition that persists past reasonable healing time (functional classification) or persists and reoccurs for more than three months (temporal classification) [124,155]. For acute pain, researchers need to differentiate between pain evoked brain responses with a no-pain state as the baseline. While for chronic pain, investigators study brain functions of pain relevant task performing or resting-state functional connectivity while the participant is under ongoing chronic pain.

To date, fNIRS imaging has been applied mostly to the acute pain condition in both clinical and experimental environments (See Figure 1). In the clinical environment, investigators have studied thermal stimuli triggered dental pain [52,54,120,131], surgery-induced pain [42,72,73,95,122,135], acupuncture evoked pin-prick pain [38,129], and similar venipuncture pain on infants [10,106,107,128,130,142]. On the other hand, in the experimental environment, researchers have studied acute pain conditions triggered by thermal [7,8,119] [13,37,51,174], mechanical [10,13,42][10,106,129,131,142][49,51,76,157,159,160], electrical [96,125] [2,35,109,110,175], chemical [40,56] stimulations, even pain imagery [170,171] and empathy [16].

The protocol designs for these stimuli can be divided into task/stimulus-based and non-task/stimulus-based designs. Task/stimulus-based design consists of event-related, block, and mixed designs [151]. In an event-related design, a single pain stimulus appears only once, or multiple pain stimuli are separated in time by an inter-stimulus interval ranging from a few to tens of seconds. Investigators usually employ the event-related design when pain stimulation is an instant process, and such stimulation often happens in the clinical environment. For example, pain caused by infant venipuncture and heel lance [10,15,23,58,106,142] and periodontal pain caused by probe insertion into labial gingival sulcus [131]. Whereas in a block design, pain conditions are presented continuously for a predefined time, ranging from a few to tens of seconds. The stimulus may be presented for a couple of times, separated by inter-block intervals, some of the examples are: thermal stimulation on the skin with a randomized inter-stimulus interval [13,174], cold pressor test [7,8,119,137,141], mechanical pressure applied to fingertip or lower back [76,159,160], repeated neuromuscular electrical stimulation [2,96,109,110,175], acupuncture operation [38,129], continuous odor presentation [40], and pain imagery by watching noxious image presentation [170,171]. Finally, there is a third type of mixed event/block design. In this type of design, a stimulus with incline/decline parameters (usually temperature or force) is applied, which will trigger pain when the participants’ sensation surpasses their pain threshold. For instance, increasing pressure on the skin [49], descending temperature applied to the hypersensitive tooth [52,54,120], or ascending temperature applied to the skin [37]. Non-task/stimulus-based designs are mostly seen in clinical studies. For instance, researchers collected regional cerebral oxygenation data during cardiac surgery periods without specific stimuli presented repeatedly [42][72]. It is worth noting that the resting-state functional connectivity (RSFC), a special type of non-task/stimulus-based design, was sometimes used to study neural networks before and/or after pain stimuli [28,52]. Also, the probes of a single fNIRS device can be divided and used for measuring two person’s hemodynamic responses simultaneously, namely hyper scanning. Such arrangement was also used in pain studies recently. For example, one study examined two participants’ brain dynamics of cooperation and competition under a capsaicin evoked acute pain condition [163].

In contrast, chronic pain requires different experimental designs to understand its etiology and pathophysiology because it is difficult to sustain a chronic pain challenge and compare it to baseline. In these pain studies, researchers usually employ various non-noxious relevant tasks or intervention procedures to test individuals with chronic pain conditions (See Figure 1). Researchers have used breath-holding tasks[4,5,28], pressure stimulation [68,157], verbal fluency task [157], [126] [29], finger tapping task [43,45], transcranial magnetic stimulation (TMS) [147] and transcranial electrical nerve stimulation (TENS) [34,36], medicine injection [164] to study patients’ brain activation under chronic pain conditions like migraine [158], fibromyalgia, chronic lower back pain, neuropathic pain, and somatoform pain disorder.

Finally, to ensure the specificity of pain evoked hemodynamic responses, studies usually involve a control group, task, or at least a rest period as a no-pain baseline. For example, itchy or low mechanical pressure stimuli accompanied with a mechanical pressure pain stimuli [76,159,160], low versus high volume current of neuromuscular electrical stimulation [2,96,109,110,175], low versus high temperatures for hand immersion in cold water or thermal stimuli [8,37,119,174], tooth percussion compared with noxious thermal dental stimulation [120], or a healthy group compared to a patient group [29,126,157].

Brain Signal Acquisition

Regions of Interest

Pain perception has a complex neurological mechanism that involves both emotional and sensory networks in the brain. Previous neuroimaging research has revealed an extensive network of brain regions associated with pain processing and perception [85]. These pain relevant brain areas include the primary and secondary somatosensory cortices (S1 and S2), the prefrontal cortex (PFC), the anterior cingulate cortex (ACC), the thalamus, and the insular cortex (IC), [80,93,152].

Due to the penetration ability of near-infrared light, fNIRS imaging can detect hemodynamic responses up to 3 cm deep, which is approximately 0.5 cm on average of the cerebral cortex according to fNIRS-fMRI signal correlation [31], and maximally 1.7 cm of brain tissue according to simulation [145], as demonstrated in Figure 2. This feature limits fNIRS to measuring only cortical brain regions. Therefore, researchers primarily focused on PFC and S1 when using fNIRS to study pain (See Figure 3). Some early fNIRS-based pain studies only placed probes over the primary sensory cortex to study the somatosensory aspect of pain [10,142], whereas a few studies collected data from only the prefrontal cortex [7,8,13,28,49,106,119], including specifically, the anterior prefrontal cortex (aPFC) [7,8,28,49,106,119], and dorsolateral prefrontal cortex (DLPFC) [13]. Thanks to the development of multi-channel fNIRS technology, later studies were able to acquire data from several different cortical regions including the prefrontal cortex, sensory cortex, premotor area, supplementary motor area, and motor cortex for pain imaging [43,52,54,72,76,96,109,110,120,157,159,160,170,171,175] (Figure 3).

Figure 2.

Different types of pain stimuli in fNIRS studies. Blue color indicates tasks/stimuli used in chronic pain studies (Image produced using GeoPain, MoxyTech, MI).

Figure 3.

Pain relevant brain areas at the cortical level investigated by fNIRS studies with animated fNIRS detection scenario.

Though the near-infrared light penetration is limited, some researchers developed methods to extend fNIRS detectability without affecting its cost-effectiveness and portability. The first method infers brain activation in the deeper region based on multi-point superficial layer activation by developed algorithms [78]. The second method augments the fNIRS signal with other portable imaging modalities like EEG [3,18]. The fNIRS-EEG combination acquires extended neurological information. However, it presents a burden on researchers for additional devices and their setting-up effort. There is no standard processing pipeline for its data merging and analysis, posing difficulties in interpreting the data. In contrast, the deep brain information inference procedure does not require additional hardware set-up. However, this is an inverse problem as we are trying to infer the deep-brain brain activation based on the hemodynamic responses detected in the superficial layer. Therefore, one has to be very careful in making such inference not to break the assumption of how light trespasses in the brain tissues. Until today, a few fNIRS-based pain imaging studies used the fNIRS-EEG combination [66,68], but none of them used the deep brain information inference method yet.

fNIRS probe localization

fNIRS provides precise localization of its probe/channel on the head surface; however, the associated anatomical information underneath is unknown. In other words, fNIRS cannot precisely locate the origins from where the functional hemodynamic responses are detected. Hence, it becomes necessary for researchers to find the relationship between the probe/channel on the head surface and underlying cortical regions. Otherwise, the detected functional responses could simply be noises measured from randomly placed fNIRS probe/channels.

There are three probe/channel registration methods for fNIRS imaging [79,156,168]. The first method is to use the 10–10/10–20 reference system, originally developed for EEG scalp electrode localization [19]. This reference system presumes a consistent relationship between scalp locations and the cortical regions beneath. This relationship has been verified using multiple structural imaging techniques, including computed tomography [50] and MRI [101]. Researchers use this validated relationship to estimate the underlying cortical region that generates the detected functional hemodynamic responses [156]. For example, the ROI of DLPFC in pain research can be localized at around F7 on the head, ROI of S1 can be localized at around C3/C4 on the homuncular head according to the 10–20 system. A software package that can estimate the probe/channel configure based on the 10–20 system is the fNIRS Optodes’ Location Decider (fOLD) toolbox [91]. However, the assumption of such a universal matching relationship sometimes brings error into group-level data analysis, given that individual-level mappings are not identical. Therefore, the same fNIRS probe/channel can measure slightly distanced cortical regions. Though there are flaws, this method is simple to implement. Therefore, most fNIRS-based pain imaging studies so far has used this registration method [28,34,36–38,40,43,52,52,54,96,109,110,126,131,157,159,160,170,171].

A second option is to precisely co-register a subject’s fNIRS data to their structural brain image scanned with MRI. This registration method merges functional hemodynamic response data with the structural MRI scan in a real-world space. Thus, this method provides higher registration accuracy due to the high-resolution MRI image. A few registration examples using MRI-based methods can be found in previous literature [27,55,79,168]. However, this method reduces the cost-effectiveness and convenience of fNIRS [156], by involving an MRI scanner in the registration process. To the best of our knowledge, no fNIRS-based pain imaging studies have used this method for probe registration yet.

A third alternative method is to use a three-dimensional (3D) digitizer (typically magnetic) or photogrammetry technology to locate fNIRS probe/channel positions with reference points in a predefined three-dimensional space [55,168]. The digitized probe/channel positions are then converted into a standard space (e.g., MNI 152 space) based on the matching reference points, respectively, in the two spaces. Finally, the probe/channel detected brain regions are estimated in the standard space, given the preloaded anatomical information. Though this method does not provide individual-level probe/channel-cortical region relationships, it accounts for individual-level differences and provides better registration results than 10–20 system-based methods [75]. Also, this method is cost-effective and portable compared with the MRI-based registration method. So far, a few fNIRS-based pain imaging studies have used this method for registration purposes [2,34,36,72,126,174,175]. A useful toolbox for this approach is the AtlasViewer toolbox [1].

Data Collection

To ensure signal quality, researchers have used spring-loaded optodes [159,160], or clear ultrasound gel [159,160] during data collection to prevent hair from staying in the light path. Additionally, fNIRS data collection is usually accompanied by physiological, clinical, and demographical data acquisition for data cleaning purposes or further joint analysis.

Physiological signals can be collected through hardware-based methods including heart rate or arterial oxygen saturation monitor [10,96,106,142,159,160], ECG [170], blood pressure [170], critical care pain observation tool [42], facial pain or motion measurement [40,42,106], the partial pressure of end-tidal carbon dioxide in arterial blood [49], respiratory rate [96,170], and skin conductance [96]. These acquired physiological signals can be used to reveal participants’ status, and help filter out physiological signals (e.g. cardiac, blood pressure, respiratory) from fNIRS signals [114].

On the other aspect, investigators collect demographic and clinical information from research participants/patients, mostly in the form of questionnaires. This type of data include visual analysis scale (VAS) or numerical rating scale (NRS) for pain levels [2,28,52,54,76,96,119,120,131,171,175], age and gender [2,7,8,28,29,37,40,42,43,49,52,72,109,110,119,120,129,157,159,160,170,171,174,175], PANAS[52,54,120], Self-Assessment Manikin scale [170], McGill [52,54,120], Pain Detect Questionnaire [159], medical history [42], handedness assessment [7,29,37,160], sleep and awake states [142], birth weight for infants [10,40], and graded chronic pain scale [157]. Some clinical studies included specific questionnaires of certain clinical conditions, like Beck Depression Inventory-II [29,157], Fibromyalgia Impact Questionnaire [28,29,43,157], Interactive Reactivity Index [157,171], hospital anxiety and depression scale [28], type of arrhythmia and number of ablations [72], locations of the hypersensitive teeth [52,120]. The collected clinical and demographic information is often analyzed jointly with the neuroimaging data to reveal associations between participants’ clinical status and pain-relevant brain dynamics.

Data preprocessing and statistical analysis

In this article, following a summarized data analysis pipeline for fNIRS in general, we also discussed the methods/steps adopted in specific fNIRS-based pain studies reviewed. The analysis of fNIRS data can be divided into three steps: preprocessing, individual-level analysis, and group-level analysis. We first discussed data conversion and filtering options in the preprocessing step. Then we reviewed the individual and group-level analysis methods in the fNIRS-based pain imaging studies.

The data is converted from the light intensity signals at two wavelengths to oxygenated and deoxygenated hemoglobin (HbO and HbR) concentration change signals [39,139]. Such conversion is done through the modified Beer-Lambert law [64,67,100]. The HbR signal has higher sensitivity at a shorter wavelength, but the shorter wavelength light has higher diffusion in the brain tissue. Quantitative analysis indicated that HbR signal changes contributed 16–22%, while HbO signal changes contributed 73–79% to the measured total cortical hemoglobin concentration changes by fNIRS [74]. Thus, many fNIRS-based pain investigations analyzed HbO signal only [34,36,37,43,72,106,109,110,120,126,131,157,171]. Whereas, a few studies calculated HbO, HbR and total hemoglobin concentration (HbT) for further statistical analysis [2,7,8,28,40,49,76,96,119,129,159,160,170,174,175].

The purpose of data cleaning is to clean the converted signal by filtering out noises and artifacts, like physiological signals (heartbeat and respiratory relevant), Mayer’s wave (blood pressure relevant), motion artifacts, and other measurement noises [20,53,114,115,148]. The denoising process can be divided into two categories: software-based and hardware-based methods. These methods were widely applied in fNIRS-based pain imaging studies. The software-based methods include bandpass filter [2,8,8,28,34,36,37,40,43,72,119,157,159,160,170,174], principal component analysis [38], independent component analysis, wavelet decomposition [34,36,37,96,120,171], interpolation [40,49,126], motion artifacts detection or removal algorithm [2,40,72,175], and data smoothing techniques including moving average [119] or Savitzky-Golay filter [159,160]. As also discussed in the data collection section, a few studies used hardware-based methods for data cleaning purposes, in which researchers employed devices like heart-rate/blood monitor, motion-sensing gyros/accelerometers, carbon dioxide pressure [49] to record physiological signals simultaneously with fNIRS data. Other studies used a specially designed short-separation channel (data channel with shorter than usual source-detector distance) to capture physiological signals and motion artifacts in the scalp [7,8,119,175] during data collection. These recorded noisy signals were then used as nuisance regressors to clean the fNIRS signal.

After the preprocessing and data cleaning steps, the first-level analysis examines the individual-level effects like brain activation and connectivity in the collected data. For brain activation, researchers usually analyze hemodynamic responses associated with noxious stimuli. The hemodynamic responses can be directly extracted by methods like minimum-maximum change value extraction [7,8,28,142], signal amplitude summation [40,49,106,131], block average [10,13,34,36,43,120,157,159,160,170,175], general linear model-based analysis [2,72,96,109,110,126,157,171,174,175], and cross-correlation analysis [76,129]. For connectivity analysis, investigators used Pearson correlation analysis [52], or wavelet coherence [38] to study the brain connectivity during the resting state. In addition, artificial intelligence algorithms like the support vector machine [37,119], K-nearest neighbor [37], and neural networks [54] were also applied at the individual level to discriminate pain from no-pain conditions.

Lastly, the group-level analysis then summarizes individual-level effects using statistical approaches when needed. Among studies reviewed in this article, researchers have applied parametric or non-parametric t-tests [2,10,28,42,72,96,120,131,142,159,160,170,175] and analysis of variance (ANOVA) [7,10,34,36,38,40,42,43,49,76,96,106,157,159,160,170] to the effects derived from the first level. Alternatively, a few studies employed mixed-effect linear models to fit and summarize the individual-level effects [8,13,109,110,126,171,174]. The most commonly used toolboxes that integrate preprocessing, individual-level, and group-level analysis are HomER [57], NIRS-Analyzer [132], and NIRS-SPM [173].

Results and Discussion

Pain evoked hemodynamic responses at the cortical level

Several studies have confirmed S1 activations contralaterally [76,125], and bilaterally [10,37,40,96,129,142,159,160,175]. Sensory cortex activation has been observed and validated in many previous studies using different imaging modalities (See meta-analysis by Apkarian 2005) [6]. Specifically, S1 plays a sensory-discriminative role in pain processing, where more blood is supplied on-site to support the neural activation evoked by the noxious stimuli [24,172]. Interestingly, two studies reported a dual-peak elevated HbO signal during thermal stimulation [13,120]. Such responses were speculated to be two responses, respectively evoked by the changing temperature and painful sensation. Also, Yucel et al. observed that bilateral HBO and HbT increases were associated with only noxious but not innocuous stimuli [175].

Another critical region reported being associated with pain, especially its emotional aspects, is the PFC, including the subdivisions of anterior PFC, DLPFC, and VLPFC [103,178]. The anterior PFC is responsible for internal state processing, memory retrieval, prospective memory, branching and reallocation of attention, and relational integration [121]. It is also known to be connected with multiple other pain relevant brain regions, including DLPFC, anterior temporal cortex, insular cortex, parietal cortex, thalamus, and basal ganglia to modulate pain perception processes [108]. Interestingly, some fNIRS based pain imaging studies reported brain activations in the PFC region [94,102,104,118], while others observed deactivations [12,92,146]. For example, PFC activations were found in studies using stimulations like cold pressor test [7,8,119], infant blood sampling processes [106], a mechanical force on the finger [76], thermal heat stimulation [13,174], neuromuscular electrical stimulation [87,96,125], and chemical smell stimulation[40]. The possible reasons for the activation include, but are not limited to, global blood volume increase dominated by the autonomic sympathetic nervous system [7,8], pain feeling suppression and disengagement [161], pain expectation and control [167], and increased communication between PFC and other cortical or subcortical regions [60,103]. Other studies that observed deactivated aPFC employed conditions like mechanical/thermal dental stimulation [120,131], electrical stimulation at the finger [1,109,110,175], cardiac surgery [72], and mechanical pressure stimulation at the lower back [49], found deactivated PFC region. The underlying mechanism of the deactivation could be the declined blood flow at PFC regions, caused by the activity suppression from the amygdala [99], the s-mPFC coordinating glutamate/GABA receptors [63], default mode network deactivation [70], vasoconstriction [69] or even non-nociceptive specific deactivation [2]. A consequential study confirmed a negative association between pain intensity and deactivation levels of anterior PFC [110]. The same research group then found deactivations in the medial anterior PFC, while there were activations in the lateral anterior PFC [109]. The DLPFC and VLPFC were known to be critically relevant to voluntary emotional regulation [178]. Also DLPFC is found to be activated in pain localization [105], and pain relevant attention (distraction) [111]. While VLPFC is reported in several social pain studies, it is more relevant to pain reappraisal strategies [47,178]. Among the articles reviewed, most studies reported activations in the DLPFC [104,120,157,169,171] and VLPFC [120,169], except one study reported deactivated DLPFC [131]. Furthermore, some other studies reported more specifically that the PFC activations might be correlated with gender, side of stimuli, depth, and handedness [106][8].

Other brain regions including the SMA, PMC, and M1, were also associated with acute pain conditions. Notably, studies found that S1, SMA, and PMA were closely related to acute pain in patients with chronic lower back pain (CLBP) and healthy participants receiving thermal stimulation on the forearms [174][159,160]. The role of SMA and PMA in pain processing is not entirely clear yet. Reasons like postural control and motion preparation under noxious conditions seem to be plausible [89,159,160]. Interestingly, a previous study confirmed that SMA played a role in pain processing [6], but its activation level also seemed relevant to pain intensity [30]. Several pain studies observed M1 activation, possibly due to the dynamic body motion or motion imagery triggered by noxious stimulation [96,125,171]. It is worth noting that several studies were able to associate brain activation strength and self-reported pain scores [8,76,125].

On another note, researchers observed abnormal hemodynamic responses in patients with various chronic pain conditions while performing different tasks, compared to healthy populations. For example, with breath-holding tasks, patients with migraine and fibromyalgia demonstrated suppressed hemodynamic responses in the prefrontal cortex [4,28]. With the verbal fluency task, patients with both somatoform pain disorder and fibromyalgia showed depressed HbO concentration increases in the prefrontal cortex [29,126]. While performing the finger-tapping task, fibromyalgia patients demonstrated depressed HbO concentration increases [34,43].

Limitations of fNIRS in pain imaging

The primary limitation of fNIRS technology is the detection depth due near-infrared light’s penetration ability, as several pain relevant brain regions are located beneath the cortex, therefore not detectable by fNIRS [65]. This limitation suggests that the optical imaging technique fNIRS may be less optimal to study the detailed brain mechanisms of pain among subcortical regions. However, pain research literature confirmed that the PFC and S1 are critical for pain processing [60,103]. Researches also reported that other cortical regions, including S1, S2, pre-motor, and supplementary motor cortices contribute to pain processes in the brain [6,24]. These cortical regions provide a window for fNIRS imaging to monitor the brain’s ongoing pain processes, particularly in the clinical environment.

Future directions for fNIRS pain imaging

Due to its portability and convenience, especially the recent development of the fibreless and wearable system [66,113,116,177], fNIRS can be used to study the performance of various complementary pain modulation techniques, like transcranial direct current stimulation (tDCS), transcranial magnetic stimulation (TMS), virtual/augmented reality (VR/AR), acupuncture, or mediation-based breathing therapy. These techniques, either need a non-ferromagnet environment or quiet operation space, which brings difficulties to patient scanning using traditional brain imaging techniques, like fMRI [44,154].

Also, with a balanced temporal and spatial resolution, fNIRS combined with the artificial intelligence (AI) algorithms provide potential means to detect patients’ pain in real-time objectively [37,54,119]. fNIRS-AI-based pain detection originates from the concept of brain-computer interface (BCI) [98]. BCI is a technology that converts signal acquired from the brain into communicative control signals, without through the regular channel of peripheral nerves and muscles [83]. In pain monitoring, the cortical signal acquired by fNIRS was analyzed using artificial intelligence algorithms to predict patients’ pain status. A list of algorithms have been tested for pain detection including support vector machine (SVM) [37,45,90,119], k-nearest neighbor (KNN) [37,45,90], linear discriminant analysis [45,51], neural networks (NN) [54].

However, the development of fNIRS-AI-based pain detection is limited by several issues at the current stage. First, the signal-to-noise ratio (SNR) of the fNIRS signal is low due to the motion artifacts and physiological noise [9,39,114,115,173]. Second, the fNIRS signal is not stationary [179]. Thus, the AI algorithms trained on collected signals sometimes fail to generalize to other data. Finally, the inter-participant variability is high, because the individualized metabolism and brain tissue structural difference often lead to different hemodynamic response baselines [88,133]. These issues brought significant challenges to feature extraction and pattern recognition when using conventional machine learning algorithms on fNIRS data [98]. One potential solution might be the recent development of deep learning (DL) and multimodal DL algorithms [46,149]. The deep learning algorithms can be trained directly on the raw fNIRS signal, as they can capture the heterogeneous features with their deep and complex structure [176]. However, these advantages brought up the requirement of a large amount of data for the DL algorithm training.

Thanks to the developing computational power, extended reality (XR) technologies, including virtual reality (VR), augmented reality (AR), and mixed reality (MR), have been widely applied nowadays in daily life [21]. These technologies created a fully immersive experience by creating a virtual environment or blending a virtual with a real environment [71]. Besides the technologies that can be used for drug-free pain modulation [48,77,143], they serve as a displaying technology in the clinical setting, for example, surgical room [33,150]. In the context of fNIRS-based pain detection, the doctors may be able to view the brain’s pain processes with the help of XR technologies. Such a prototype framework has been preliminarily tested by our group in a feasibility study, named clinical augmented reality and artificial intelligence (CLARAi) [54]. Imagine the scenario in a dental clinic in the future, that the dentist is wearing an AR glass while performing oral surgery, through which he/she can objectively see the patient’s pain status and its relevant brain processes so that the patient doesn not have to speak out when he or she feels pain.

Conclusion

This paper has reviewed state-of-the-art fNIRS technology applied in pain neuroimaging by discussing all components in fNIRS neuroimaging. Starting from experimental design, we discussed several types of stimulation that evoke acute pain and tasks that investigators employed to study chronic pain. We then reviewed the brain regions that were monitored in studies, including prefrontal, sensory, motor, pre-motor, and supplementary cortices. Also, we discussed a crucial step of probe localization before fNIRS data collection. Then, we categorized methods in different fNIRS data analysis steps, including data preprocessing, individual-level, and group-level analysis. Finally, we briefly summarized the findings revealed by fNIRS-based pain neuroimaging studies, then discussed the limitation of the technology and its future in the field.

We concluded that there are many avenues for future fNIRS-based pain neuroimaging research, particularly in its applications in the clinical environment, accompanying non-invasive pain neuromodulation methods, AI, and XR technologies. Although fNIRS-AI pain evaluation applications have been demonstrated in several studies, no commercial fNIRS-AI application is currently available. However, the research trends predict that the fNIRS-AI application will continue to grow. In the near future, breakthroughs via fNIRS-AI pain imaging systems are expected to be seen in the clinical environment.