Dynamic Quantification of Migrainous Thermal Facial ... · PAVLIDIS et al.: DYNAMIC QUANTIFICATION OF MIGRAINOUS THERMAL FACIAL PATTERNS – A PILOT STUDY 1227 Fig. 1. Middle Row:

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 23, NO. 3, MAY 2019 1225

Dynamic Quantification of Migrainous ThermalFacial Patterns – A Pilot Study

Ioannis Pavlidis , Senior Member, IEEE, Ivan Garza , Panagiotis Tsiamyrtzis, Malcolm Dcosta,Jerry W. Swanson, Thomas Krouskop , and James A. Levine

Abstract—This article documents thermophysiologicalpatterns associated with migraine episodes, where the in-ner canthi and supraorbital temperatures drop significantlycompared to normal conditions. These temperature dropsare likely due to vasoconstriction of the ophthalmic arter-ies under the inner canthi and sympathetic activation ofthe eccrine glands in the supraorbital region, respectively.The thermal patterns were observed on eight migraine pa-tients and meticulously quantified using advance compu-tational methods, capable of delineating small anatomicalstructures in thermal imagery and tracking them automati-cally over time. These methods open the way for monitor-ing migraine episodes in nonclinical environments, wherethe patient maintains directional attention, such as his/hercomputer at home or at work. This development has the po-tential to significantly expand the operational envelope ofmigraine studies.

Index Terms—Migraine, headache, thermal imaging, facetracking, facial features, supraorbital, periorbital, maximumlikelihood estimation.

I. INTRODUCTION

THE diagnosis of migraine headache is performed clini-cally and relies heavily on specific criteria outlined in the

International Classification of Headache Disorders, 3rd Edition(ICHD-3) [1]. Screening tools, such as the ID Migraine, facili-tate diagnosis by improving migraine recognition [2]. Children’sdrawings have also proven to aid in the diagnosis of migraine inthis population [3]. All currently available diagnostic methods,however, require effective communication between the exam-iner and patient, as an accurate migraine diagnosis is not based

Manuscript received January 31, 2018; revised June 3, 2018; acceptedJuly 10, 2018. Date of publication July 12, 2018; date of current versionMay 6, 2019. This work was supported in part by the National ScienceFoundation under Grant IIS-0414754, in part by the Texas Medical Cen-ter, in part by the Methodist Hospital, and in part by the Mayo ClinicFoundation for Research. (Corresponding author: Ioannis Pavlidis.)

I. Pavlidis is with the Computational Physiology Lab, University ofHouston, Houston, TX 77204 USA (e-mail:, [email protected]).

I. Garza and J. W. Swanson are with the Mayo Clinic, Rochester, MN55905 USA (e-mail:,[email protected]; [email protected]).

P. Tsiamyrtzis is with the Department of Statistics, Athens University ofEconomics and Business, 10434 Athens, Greece (e-mail:,[email protected]).

M. Dcosta is with the Department of Mathematics and Computer Scie-nce, Elizabeth City State University, Elizabeth City, NC 27909 USA(e-mail:,[email protected]).

T. Krouskop is with the Baylor College of Medicine, Houston, TX 77030USA (e-mail:, [email protected]).

J. A. Levine is with the Fondation IPSEN, 92650 Paris, France (e-mail:,[email protected]).

Digital Object Identifier 10.1109/JBHI.2018.2855670

on pain presence alone, but also on the presence or absenceof migraine-associated symptoms such as photo- and phono-phobia, nausea, or aura. Available scales relying on observationof pain behaviors can assist pain assessment in those with signif-icant cognitive impairment and limited communication ability[4], but cannot assess migraine-associated symptomatology.

Significant advances have recently been made in understand-ing the pathophysiology of migraine headache, in part due toneuroimaging techniques using functional Magnetic ResonanceImaging (fMRI) and Positron Emission Tomography (PET) [5],[6]. These modalities allow observation of distinct cerebral re-gions (e.g., dorsal pons) and networks, as they become activatedand involved during a migraine headache attack [7]. Migraine’smechanisms, however, remain incompletely understood [8]. Itsepisodic nature creates challenging logistic problems when at-tempting to study spontaneous headache events. A conclusiondrawn from these studies that is of practical significance is thattreating the patient with triptans early in the migraine attack(before the development of cutaneous allodynia), increases thelikelihood of pain-freedom [9]. We are interested in develop-ing a field (vs. a clinical) method for the quantitative study andpossibly detection of migraines. We have focused on thermalimaging of the face, because the face is most often exposed indaily living conditions and thermal imaging is passive and hencesafe for prolonged monitoring [10]. Our investigation aimed tocontribute towards three open problems in the study and treat-ment of migraines:

� Facilitate the investigation of the episodic nature of thecondition. The thermal imaging sensor can be attached asa peripheral on a personal computer, enabling monitoringduring work hours or at home. Such monitoring can covera substantial portion of the day and thus, it would signifi-cantly increase the chances of capturing the developmentof migraine attacks. To make such monitoring feasible,however, the thermal imaging system should be capableof tracking facial features of interest in the presence ofnatural head motion.

� Localize with precision migrainous thermal patterns in fa-cial areas of neurophysiological importance, aiding repro-ducibility and the understanding of underlying processes.

� Bridge the communication gap between the patient andthe clinician. Assuming that thermal facial patterns arefound to be associated with migraines, the method canbe especially useful as a diagnostic aid for patients withlimited communication abilities.

2168-2194 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

1226 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 23, NO. 3, MAY 2019

Consequently, the central research question is how exactlymigraine headache attacks affecting facial thermophysiologicalresponses and what methods to use in order to capture suchresponses dynamically. To answer this question we conducted anobservational study, with methods and results that are presentedin the remainder of the paper.

This is not the first attempt to investigate the diagnostic poten-tial of thermal imaging in migraine and other types of headaches.Considerable interest in migraine evaluation using thermogra-phy took place in the 1970s and 1980s hoping to better under-stand the disorder and use it as a diagnostic and therapeutictool [11]. Some studies, however, failed to find specific thermo-graphic patterns. Wood documented a pattern of cooling seenin one supraorbital region and periorbital areas in 63% of clus-ter headache patients, but this only rarely occurred in migraine[12]. Others subsequently reproduced very similar results [13].

Other investigators found evidence suggesting specific ther-mographic patterns could be seen during a migraine headacheattack and normalize interictally. Such findings, however, werenot uniform. Drummond reported a higher average orbital tem-perature of 35.6 ◦C during a migraine headache and 35.4 ◦Cwhen headache free [14]. Higher temperatures during migraineheadache vs. the asymptomatic period when using the angularorbit and supraorbital arterial reference points have also beendocumented [15].

In contradistinction to these results, multiple other studiessuggested facial cooling during migraine might be a more spe-cific pattern seen during acute migraine. Lance reported fore-head cooling in 8 of 12 subjects during spontaneous hemicranialmigraine headache ipsilateral to the pain, the temperature pro-gressively dropped as pain intensified [16]. Similarly, a coldpatch in the external carotid territory ipsilateral to the prevailingside of pain was seen in 13 out of 17 subjects with migrainewith and without aura in Dalla Volta’s study. The cold patchdisappeared or attenuated in parallel with clinical improvementfollowing treatment [17]. Subsequently, the same group repli-cated and extended the prior findings now with a larger sampleof 246 migraine patients. Of these, 206 exhibited the typical coldpatch in the forehead. Among the 136 patients who experiencedcomplete or substantial relief from headache the cold patch dis-appeared or markedly improved in 85% of the cases. The authorssuggested thermography could be useful to monitor the clinicalcourse of the disease and could represent a useful criterion forthe decision of discontinuing preventive therapy [18]. Swerd-low and Dieter [19] disputed some of these findings, arguingthat the ‘vascular cold patch is independent of prognosis andis most likely a permanent element of a vascular headache suf-ferer’s facial thermal pattern.’ A possible explanation for thesecontradicting research outcomes lies in the non-uniform criteriaused to select patients - a factor that needs to be taken seriouslyin migraine studies.

Our migraine study differs from prior efforts with respect toits design, analysis, and technical methods. In terms of design,we used strict patient screening criteria, collecting data not onlyinterictally, but also episodically. In terms of analysis, we per-formed investigation both at the individual and group levels.In terms of technical methods, we introduced computational

algorithms that operate on thermal imaging sequences ratherthan still thermal images, affording continuous measurement onanatomical markers despite small head motions. In the presentstudy, applying these methods to data collected via a highdefinition thermal imaging sensor, we clarified the underlyingphysiological mechanisms in the periorbital and supraorbitalregions, during and after a migraine attack. Importantly, oursemi-automated methods can potentially be used in free-livingscenarios featuring directional attention, such as computer workat the office - a development that could significantly expand theoperational envelope of migraine studies.

II. METHODS

A. Approvals

Approvals were obtained from the Institutional ReviewBoards of Mayo Clinic and University of Houston. Each subjectunderwent an informed consent process and provided writteninformed consent prior to participation.

B. Subject Inclusion and Exclusion Criteria

We searched for patients seen for migraine in the Departmentof Neurology at Mayo Clinic, Rochester MN. Using medicalrecord retrieval, we identified 1576 subjects. Of these, we iden-tified 70 as having episodic migraine on no preventive medica-tion and living in the vicinity of Mayo Clinic to facilitate thestudy. We sent a letter to them inviting them to participate andof the 70 patients, 31 responded with interest. We interviewedall 31 subjects to confirm the diagnosis and review inclusionand exclusion criteria. Of the 31 candidates, we excluded 13because they had other coexistent headache disorders, did nothave episodic migraine, were overusing acute treatments, orwere using migraine preventives.

C. Collection of Subject Data

We instructed all 18 subjects who were included in the studyto come to the laboratory for thermal video recording at thebeginning of a migraine headache. We also instructed them tonot treat the headache before the recording. In such visits, anexaminer from our clinic determined if the subject had indeedan episode of migraine headache with or without aura accordingto the International Classification of Headache Disorders, 3rdEdition (ICHD-3) [1]. Then, the study coordinator recorded athermal video of the subject’s face for 5 min (‘Migraine’ ses-sion). At that time, the subject was free to treat her headache andleave the lab. We asked the subject to return when her symptomshad stopped and the headache was gone. In this second visit,the study coordinator recorded again a thermal video of the sub-ject’s face for 5 min (‘Baseline session). We obtained from allsubjects their age, sex, and medication history.

D. Imaging Parameters

We used a mid-wave infrared camera (FLIR SC 4000 FLIRSystems, Boston, MA, USA) for thermal video recording. Thecamera featured an InSb focal plane array of 320 × 256 pixels

PAVLIDIS et al.: DYNAMIC QUANTIFICATION OF MIGRAINOUS THERMAL FACIAL PATTERNS – A PILOT STUDY 1227

Fig. 1. Middle Row: Supraorbital and periorbital tissue trackers at work as subject D006 exhibits small head motion during the baseline session.Top Row: Motion-corrected periorbital snapshots with the segmented thermal footprints of the ophthalmic arteries - the corresponding signal pointsare indexed above. Bottom Row: Motion-corrected supraorbital snapshots with the corresponding signal points indexed below.

and thermal sensitivity of 0.025 ◦C; the recording speed was setat 54 fps. We focused on the supraorbital and periorbital areastwo facial regions of neurophysiological importance, for whichother researchers reported migrainous thermal patterns, but in arather qualitative, non-algorithmic manner (e.g., ‘cold patch in[17], [18]). From each thermal video we extracted the followingthermophysiological responses:

1) Supraorbital thermal signal S2) Periorbital thermal signal P

Each signal point represented the mean temperature in ◦C inthe respective area of the face at a particular time.

As the signals S and P were extracted from imaging mea-surements with no restrictions imposed on the subjects, otherthan sitting on a chair, we needed methods for virtual prob-ing and tethering. These are the computational equivalents forthe actual probe and its attachment paraphernalia in traditionalclinical measurements (e.g., thermistor). In medical imaging ter-minology virtual tethering is called tissue tracking, while virtualprobing is called tissue segmentation. Tracking and segmenta-tion are important to the envisioned practical application of thismethodology, as subjects who work in front of personal com-puters frequently exhibit small motion and have differing facialcharacteristics.

E. Thermal Imaging - Tissue Tracking

The virtual tissue tracker encompassed and tracked of thesegmented region of interest, despite motions of the subject’sface (Fig. 1). This ensured that the thermophysiological signalextractors operated on consistent and valid sets of data overthe data collection timeline. We used the tissue tracking algo-rithm we reported in [20]. Other tracking algorithms exist inthe literature [21], and can potentially be employed in migrainestudies. We chose the specific tracker because it was validated

for precision tracking of facial tissue in thermal imaging [20],and subsequently was used with success by multiple groups inchallenging studies [22], [23]. Hence, it was an appropriate,meritorious, and safe choice.

The tracker features a particle filter, driven by a probabilistictemplate function with both spatial and temporal smoothingcomponents, which is capable of adapting to abrupt positionaland physiological changes. On the initial frame of each thermalsequence in our dataset, the operator initiated two such trackersby selecting via mouse the subject’s supraorbital and orbitalregions as follows:

Localization of Supraorbital Tracker. The supraorbital trackerwas a rectangle with a base that bridged the inner ends of thesubject’s eyebrows and extended halfway up the height of theforehead (Fig. 1). In fact, the main part of the tracker coversthe supratochlear region, and could have been called ‘supra-tochlear tracker’. However, because it marginally extends tothe broader supraorbital area, and because such trackers arereferred as ‘supraorbital trackers’ in the technical literature[24], we chose to keep the convention, while making thisclarifying note.

Localization of Periorbital Tracker. The periorbital trackerwas a rectangle that included the two orbits height-wise, butleft out their outer halves width-wise (Fig. 1).

After tracker selection in the initial frame of the thermalclip, the rest of the computational process was automated. Eachtracker estimated the best matching block for the next frame ofthe thermal clip based on particle filtering driven by spatiotem-poral smoothing. Particle filtering can handle nonlinear motion,which head motion mostly is, while spatiotemporal smoothingdoes away with the unrealistic assumptions of pixel and frameindependence, thus, increasing accuracy and reducing trackingoscillation.

1228 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 23, NO. 3, MAY 2019

TABLE IEXPERIMENTAL VARIABLES

F. Thermal Imaging - Segmentation & Signal Extraction

The supraorbital and periorbital signals were extracted bycomputing the mean temperature of the corresponding anatom-ical landmarks in each time step. These anatomical landmarkswere segmented within the supraorbital and periorbital trackers,respectively.

Segmentation of Supraorbital Landmark. For the supraor-bital region the signal was extracted by computing in eachtime step the mean temperature in the entire tracking area;therefore, segmentation of the supraorbital landmark was triv-ial (Fig. 1). This landmark was chosen because it includes thesupraorbital arteries and a preponderance of eccrine glandsboth of neurophysiological importance.

Segmentation of Periorbital Landmark. For the periorbitalregion the signal was extracted by computing in each timestep the mean temperature in the thermal footprints of the leftand right ophthalmic arteries - a small portion of the over-all tracking area located in the inner canthi. A segmentationalgorithm, operating within the periorbital tracker, was de-lineating the arteries’ apparent footprints in each incomingframe (Fig. 1). We reported the details of this algorithm andits operational characteristics in [25]. In every frame the algo-rithm delineated the region of interest in each orbit by startingfrom the local maximum (seed) and expanding according toa probabilistic cost function. This cost function factored ingeometry, temperature homogeneity, and temperature gradi-ent space adjacency. The local maximum corresponded to thehottest local pixel, which based on heat transfer laws wasbound to be in the center of arterial blood flow. This land-mark was chosen because the ophthalmic arteries anchor neu-rophysiological responses in the orbital area, as they supplywith blood the ocular muscles.

G. Statistical Analysis

We sought to find if there were significant differences in thethermophysiological responses between the baseline and the

migraine sessions. Because we used two tests (supraorbital tem-perature comparison and periorbital temperature comparison)we applied Bonferroni correction for the level of significance(α = 0.05/2 = 0.025).

We adopted a two-stage analysis. In the first stage we ex-amined what happened in each subject. In the second stage wetested if any significant trends observed in individuals held forthe entire group of subjects. Table I lists all the experimentalvariables used in the analysis. Please note that we treated thepain as a binary entity present in the migraine session and absentin the baseline session).

The mean supraorbital and periorbital temperatures for themigraine and baseline sessions ( Sm ,Sb ,Pm ,Pb , respec-tively) were computed by averaging over 16,200 measurements(54 fps × 60 s × 5 min per session) a highly dense temporalsupport of non-trivial duration.

III. RESULTS

A. Subject Characteristics

At the time the study ended, eight subjects completed thestudy and were recorded while 10 did not have a migraine at-tack to study. None of the completers had prior or active au-tonomic disorders. All completers (n = 8) were female; sevenhad episodic migraine without aura and one episodic migrainewith aura per ICHD-3. Age range was 21–57 years old, withmean ± SD = 37 ± 11.4 years old.

B. Supraorbital Temperatures in Baseline vs. Migraine -Subject Level

Every subject had significantly lower mean supraorbital tem-perature Sm in the migraine session with respect to the meansupraorbital temperature Sb in the baseline session (p < 0.0001,two-sample t-test for all). Figure 2 shows detailed resultsfor a representative case (Subject D003), while Figure 3shows the supraorbital temperature distributions for the entiresubject set.

PAVLIDIS et al.: DYNAMIC QUANTIFICATION OF MIGRAINOUS THERMAL FACIAL PATTERNS – A PILOT STUDY 1229

Fig. 2. (a) Facial thermal image of subject D003 representative of her baseline condition and (b) of her migraine condition. (c) Measurementareas. (d) Comparative boxplots. Each box plot contains the thermal signal values in the specic area (supraorbital or periorbital) and for the speciccondition (baseline or migraine). The mean values are indicated by the ‘*’ symbol.

C. Periorbital Temperatures in Baseline versusMigraine - Subject Level

In 7 out of the 8 subjects the migraine session had significantlylower mean periorbital temperature Pm with respect to the pe-riorbital temperature Pb in the baseline session (p < 0.0001,two-sample t-test); for one subject (D009) there was no signifi-cant difference (p = 0.0458, two-sample t-test). Figure 2 showsdetailed results for a representative case (Subject D003), while

Figure 4 shows the periorbital temperature distributions for theentire subject set.

D. Hypothesis Validity across Subjects

From the analysis thus far we have inferred that during mi-graine sessions the subjects have lower supraorbital and perior-bital temperatures with respect to their baseline sessions. Thisis true for all eight subjects regarding the supraorbital response

1230 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 23, NO. 3, MAY 2019

Fig. 3. Boxplots of the supraorbital thermal signals in the baseline and migraine sessions for each subject. We performed a two-sample t-test foreach subject, to decide if the mean supraorbital response variable has significant differences between the baseline and migraine sessions. All thep-values are far less than α = 0.025 (p < 0.0001). The mean values are indicated by the ‘*’ symbol.

and for seven out of the eight subjects regarding the periorbitalresponse. While these intra-individual results are robust, be-cause they are based on thousands of thermal measurements persubject, a legitimate concern is if they will scale up in a largersubject sample.

To address this concern, we test whether this finding consti-tutes a significant (at α = 0.025) result or not at the group level.In other words, we seek to find how far from randomness thisresult is, and thus, how likely it is to be repeated in a futurestudy. We employ a Bayesian methodology to bypass issueswith small sample sizes, using an agnostic prior to be conserva-tive. Specifically, for each subject i = 1,2, . . . , 8 we define therandom variables Xi and Yi as:

Xi =

{0, if Sb ≤ Sm

1, if Sb > Sm

and Yi =

{0, if Pb ≤ Pm

1, if Pb > Pm

(1)

The random variable Xi (Yi) is 1 if the research hypothe-sis is true for the supraorbital (periorbital) temperatures and 0otherwise. Each of the Xi and Yi forms a Bernoulli randomvariable with probabilities of success (assumed constant acrosssubjects):

θ = Pr(Sb > Sm ) and φ = Pr(Pb > Pm ) (2)

If we assume that the subjects are independent of each other,then the sum of the random variables Xi over the eight subjectsforms a binomial distribution and so does the sum of the randomvariables Yi :

X =8∑

i=1

Xi ∼ Bin(8, θ) and Y =8∑

i=1

Yi ∼ Bin(8, φ) (3)

In our data set we get X = 8 and Y = 7. Our interest lies inthe probabilities of success θ and φ. Specically, we are inter-ested to determine if these probabilities are significantly higherthan 0.5, which represents random guess. We have a relativelysmall number of subjects in the study (n = 8) and the pointestimates in both experiments are near or at the edge of theparameter space [0, 1] (the Maximum Likelihood Estimates ofθ, φ are θ̂ = 8/8 = 1 and φ̂ = 7/8 = 0.875). Hence the standardfrequentists methods relying on the asymptotic performance arequestionable; thus, we opted for a Bayesian analysis.

We have two Binomial experiments with likelihoods:

X|θ ∼ Bin(8, θ) and Y |φ ∼ Bin(8, φ), (4)

and we are interested in drawing an inference about the successprobabilities θ and φ. Initially we need to provide prior distri-butions for θ and φ. Since no prior knowledge is available to us,

PAVLIDIS et al.: DYNAMIC QUANTIFICATION OF MIGRAINOUS THERMAL FACIAL PATTERNS – A PILOT STUDY 1231

Fig. 4. Boxplots of the periorbital thermal signals in the baseline and migraine sessions for each subject. We performed a two-sample t-test foreach subject, to decide if the mean periorbital response variable has significant differences between the baseline and migraine sessions. All thep-values but one are far less than α = 0.025 (p < 0.0001 for all but subject D009, for which p = 0.0458). The mean values are indicated by the ‘*’symbol.

we adopt the Uniform distribution in the range [0, 1] for bothparameters:

π(θ) ∼ U [0, 1] and π(φ) ∼ U [0, 1] (5)

Bayes theorem provides the posterior distribution for each pa-rameter given the observed data. Specifically, it is easy to showthat:

θ|X = 8 ∼ Beta(9, 1) and φ|Y = 7 ∼ Beta(8, 2). (6)

The distribution plots in Figure 5 show that the posterior prob-ability mass of the parameters θ and φ lies to the far right andaway from 0.5. This can be quantified by computing the HighestPosterior Density (HPD) interval for each parameter when theconfidence level is at 0.025, to match the Bonferroni correctedvalue in the earlier tests:

97.5% HPD interval for θ is [0.664, 1.000]

97.5% HPD interval for φ is [0.515, 0.995].

In neither case the value of 0.5 is included, which means thatthe random guess scenario is excluded and our hypothesis isaccepted - thus, the pattern that was observed in these eightsubjects is likely to be reproduced in future trials.

IV. DISCUSSION

There are significant temperature decreases in the supraorbitaland periorbital areas during a migraine attack - a pattern that ouranalysis shows is likely to be widespread. The mechanism andrationale for this reduction are not completely understood. Butin the case of the supraorbital region the temperature reductionappears to be partly due to the onset of perspiration, as evidencedby the blobby thermal pattern. And in the case of the periorbitalregion the reduction appears to be due to vasoconstriction of theophthalmic arteries, the thermal footprints of which are centerednext to the lacrimal ducts. Should this etiology of the tempera-ture reduction be correct, then both responses are likely to be ofsympathetic origin. This fresh insight is afforded in part by thenew generation high definition thermal imaging sensors and inpart by the computational methods we used.

Irrespective of the phenomenon’s origin, migraines appear tobe associated with a characteristic facial thermal pattern, whichcan be detected based on simple t-tests between the evolvingmean temperature distributions in the supraorbital and perior-bital areas and their baseline values. Real-time measurementsand comparisons can be performed via a thermal imaging sys-tem mounted atop a personal computer. Importantly, the tis-sue areas where these measurements are performed are defined

1232 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 23, NO. 3, MAY 2019

Fig. 5. Plots of the prior and posterior distributions for the supraorbital and periorbital random experiments.

and tracked algorithmically, facilitating reproducibility andquantification.

We imposed strict criteria for subject selection to control con-founding variables. We also aimed to record diagnosed migraineepisodes in real-time. This ambitious design in combinationwith the state-of-the-art computational methods produced highquality data in this study. At the same time, however, the studydesign resulted in a small set of subjects. To address issues in-herent in small subject sets, we performed a two-level analysis- a frequentist inference at the intra-individual level (where wehad thousands of measurements and we could afford it), and aBayesian inference at the group level (where we had few datapoints and we could not afford a frequentist approach). Theresults showed with high degree of confidence that the strongthermophysiological pattern observed within individuals is non-random, and thus likely to scale up to larger samples.

A legitimate question is if this thermophysiological pattern onthe face characterizes other activities in which subjects engagewhen they are in front of a computer. In a series of past studies wedocumented characteristic facial thermophysiological patternsassociated with common desktop activities - none appears to bein conflict with the observed migraine pattern. Specifically, wedocumented that cognitive loading is associated with gradualwarming of the supraorbital area [26], startle is associated withinstantaneous warming of the periorbital area [27], and chewingis associated with gradual warming of the mandible area [28].One unsettled issue is if headaches other than migraines producethe same facial themophysiological signature - something thatcalls for further research.

There are some practical limitations regarding the applica-tion of the method. The periorbital area is not accessible insubjects who wear glasses. There is, however, redundancy inthe methodology and in this case monitoring could be based onthe supraorbital signal only. The supraorbital area itself may beinaccessible in subjects with hair banks, but this problem caneasily be solved with bobby pins.

We note that all the subjects were female. This was not adeliberate study choice but reflects the fact that migraines aremore prevalent in females than in the male population [29],[30]. It is likely that male migraine sufferers manifest the same

thermophysiological pattern on the face. However, because therewere no male subjects in our experimental set, this remains aquestion for a future study.

A method such as the one described in this study, could pro-vide observational access in migraine studies outside the clin-ical setting and facilitate gaining a better understanding of theepisodic nature of the ailment. Migraines are not fully under-stood, partly because it has been difficult to study them duringdaily living activities. The problems associated with studyingmigraines are exacerbated when patients have communicationdifficulties. The methodology described in this paper may pro-vide a useful means of conducting field studies of migrainepatients and better understanding the etiology of the attacks.

V. CONCLUSION

Migraine episodes appear to be accompanied by temperaturereduction in the periorbital and supraorbital areas. These phe-nomena could be detected by computational methods operatingon thermal imagery. The rapid reduction in size and price ofhigh quality thermal imaging sensors, and the commoditizationof cloud computing, bring the results of this study within strik-ing distance of practical applications. Such applications wouldenable the study of migraines ‘in the wild’, and optimize the ad-ministration of acute migraine treatment. Before attempting totranslate this research into clinical and field practice, however,studies with larger subject sizes and investigation of comorbidityfactors are in order.

ACKNOWLEDGMENT

The authors would like to thank L. MacBride for her helpin collecting data for this study. This material is based uponwork supported by the National Science Foundation award IIS-0414754 entitled “Interacting with Human Physiology.” It wasalso supported in part by grants from the Texas Medical Center,the Methodist Hospital, and the Mayo Clinic Foundation forResearch. Any opinions, findings, and conclusions or recom-mendations expressed in this paper are those of the authors anddo not necessarily reflect the views of the funding agency.

PAVLIDIS et al.: DYNAMIC QUANTIFICATION OF MIGRAINOUS THERMAL FACIAL PATTERNS – A PILOT STUDY 1233

REFERENCES

[1] Headache Classification Subcommittee of the International Headache So-ciety, “The international classification of headache disorders, 3rd edition,”Cephalalgia, vol. 33, no. 9, pp. 629–808, 2013.

[2] R. Lipton et al., “A self-administered screener for migraine in pri-mary care: The ID Migraine validation study,” Neurology, vol. 61, no. 3,pp. 375–382, 2003.

[3] C. Stafstrom, K. Rostasy, and A. Minster, “The usefulness of children’sdrawings in the diagnosis of headache,” Pediatrics, vol. 109, no. 3,pp. 460–472, 2002.

[4] R. van Herk, M. van Dijk, F. Baar, D. Tibboel, and R. de Wit, “Observationscales for pain assessment in older adults with cognitive impairments orcommunication difficulties,” Nursing Res., vol. 56, no. 1, pp. 34–43, 2007.

[5] T. Schwedt and D. Dodick, “Advanced neuroimaging of migraine,” LancetNeural, vol. 8, pp. 560–568, 2009.

[6] A. May and M. Matharu, “New insights into migraine: Application offunctional and structural imaging,” Current Opinion Neurology, vol. 20,no. 3, pp. 306–309, 2007.

[7] S. Afridi et al., “A positron emission tomographic study in spontaneousmigraine,” Archives Neurology, vol. 62, no. 8, pp. 1270–1275, 2005.

[8] D. Pietrobon and J. Striessnig, “Neurobiology of migraine,” Nature Rev.Neuroscience, vol. 4, pp. 386–398, 2003.

[9] R. Burstein, B. Collins, and M. Jakubowski, “Defeating migraine painwith triptans: A race against the development of cutaneous allodynia,”Ann. Neurology, vol. 55, no. 1, pp. 19–26, 2004.

[10] I. Pavlidis, J. Dowdall, N. Sun, C. Puri, J. Fei, and M. Garbey, “Interactingwith human physiology,” Comput. Vision Image Understanding, vol. 108,no. 1–2, pp. 150–170, 2007.

[11] K. Ammer, “Does thermal imaging provide extra information in patientssuffering from headache?” Thermology Int., vol. 16, no. 2, pp. 45–48,2006.

[12] E. Wood, A. Friedman, and L. Holmes, “Thermography in migraine,”Trans. Amer. Neurological Assoc., vol. 96, pp. 192–194, 1971.

[13] A. Friedman, “Migraine: Variations on a theme,” The Fourth Cyril B.Courville Lecture - Bulletin Los Angeles Neurological Societies, vol. 40,no. 3, pp. 83–95, 1975.

[14] P. Drummond and J. Lance, “Facial temperature in migraine, tension-vascular and tension headache,” Cephalalgia, vol. 4, no. 3, pp. 149–158,1984.

[15] M. Giacovazzo, P. Martelletti, and G. Valducci, “Computerized telether-mographic assessment in migraine, with particular reference to the pro-dromal phase,” Cephalalgia, vol. 6, no. 4, pp. 219–222, 1986.

[16] J. Lance and M. Anthony, “Thermographic studies in vascular headache,”The Med. J. Aust., vol. 1, no. 5, pp. 240–243, 1971.

[17] G. D. Volta and G. Anzola, “Are there objective criteria to follow upmigrainous patients? A prospective study with thermography and evokedpotentials,” Headache, vol. 28, no. 6, pp. 423–425, 1988.

[18] G. D. Volta, G. Anzola, and V. DiMonda, “The disappearance of the‘cold patch’ in recovered migraine patients: Thermographic findings,”Headache, vol. 31, no. 5, pp. 305–309, 1991.

[19] B. Swerdlow and J. N. Dieter, “The vascular ‘cold patch’ is not a prognosticindex for headache,” Headache: The J. Head Face Pain, vol. 29, no. 9,pp. 562–568, 1989.

[20] Y. Zhou, P. Tsiamyrtzis, P. Lindner, I. Timofeyev, and I. Pavlidis, “Spa-tiotemporal smoothing as a basis for facial tissue tracking in thermalimaging,” IEEE Trans. Biomed. Eng., vol. 60, no. 5, pp. 1280–1289,May 2013.

[21] N. Sharma, A. Dhall, T. Gedeon, and R. Goecke, “Modeling stress usingthermal facial patterns: A spatio-temporal approach,” in Proc. HumaineAssoc. Conf. Affective Comput. Intell. Interact., 2013, pp. 387–392.

[22] G. Ponsi, M. S. Panasiti, G. Rizza, and S. M. Aglioti, “Thermal facial reac-tivity patterns predict social categorization bias triggered by unconsciousand conscious emotional stimuli,” Proc. Roy. Soc. B, vol. 284, no. 1861,Art. no. 20170908, 2017.

[23] I. Pavlidis et al., “Dissecting driver behaviors under cognitive, emotional,sensorimotor, and mixed stressors,” Scientific Rep., vol. 6, Art. no. 25651,2016.

[24] D. Shastri, A. Merla, P. Tsiamyrtzis, and I. Pavlidis, “Imaging facial signsof neurophysiological responses,” IEEE Trans. Biomed. Eng., vol. 56,no. 2, pp. 477–484, Feb. 2009.

[25] D. Shastri, P. Tsiamyrtzis, and I. Pavlidis, “Periorbital thermal signalextraction and applications,” in Proc. 30th Int. IEEE Conf. Eng. Med.Biol., 2008, pp. 102–105.

[26] J. Levine, I. Pavlidis, L. MacBride, Z. Zhu, and P. Tsiamyrtzis, “De-scription and clinical studies of a device for the instantaneous detectionof office-place stress,” WORK: A J. Prevention, Assessment, Rehabil.,vol. 34, no. 3, pp. 359–364, 2009.

[27] J. Levine, I. Pavlidis, and M. Cooper, “The face of fear,” The Lancet,vol. 357, no. 9270, p. 1757, 2001.

[28] I. Pavlidis, J. Levine, and P. Baukol, “Thermal imaging for anxiety de-tection,” in Proc. IEEE Workshop Comput. Vision Beyond Visible Spectr.:Methods Appl., Jun. 2000, pp. 104–109.

[29] R. Lipton, W. Stewart, S. Diamond, M. Diamond, and M. Reed, “Preva-lence and burden of migraine in the United States: Data from theAmerican Migraine Study II,” Headache, vol. 41, no. 7, pp. 646–657,2001.

[30] R. B. Lipton et al., “Migraine prevalence, disease burden, and theneed for preventive therapy,” Neurology, vol. 68, no. 5, pp. 343–349,2007.