Microvascular Research 68 (2004) 104 – 109

www.elsevier.com/locate/ymvre

Analysis of IR thermal imager for mass blind fever screening

Eddie Y.K. Ng, a,* G.J.L. Kaw, b and W.M. Chang c
a
School of Mechanical and Production Engineering Nanyang Tech. University, Singapore 639798, Singapore
b
Department of Diagnostic Radiology, Tan Tock Seng Hospital, Singapore 308433, Singapore
c
Zugo Photonics Pte Ltd, Kaki Bukit Techpark II, Singapore 415976, Singapore
Received 27 February 2004
Available online 1 July 2004

Abstract

Background: Obtaining meaningful temperature for the human body requires identifying a body site that will provide reliable data across a
large population. It is important to understand that skin temperature does not solely depend on body-core temperature and may be affected by
other physiological and environmental factors. Currently, there is lack of empirical data in correlating facial surface temperature with body
core temperature. Present IR systems in use at airports/immigration checkpoints have not been scientifically validated particularly in regards
to the false-negative rate. As a result, they may create a false sense of security by underestimating the number of febrile (and possibly
infected) individuals. This article evaluates the effectiveness of thermal scanner when it is being used for mass blind screening of potential
fever subjects such as SARS or bird flu patients. Methods: Bio-statistics with regression analysis and ROC is applied to analyse the data
collected (502) from the SARS hospital in Singapore and conclusive results are drawn from them. The results are vital in determining two
very important pieces of information: the best and yet practical region on the face to take readings and optimal pre-set threshold temperature
for the thermal imager. Results: (1) The thermal scanner can be used as a first line tool for the mass blind screening of hyperthermia, (2) the
readings from the scanner suggest good correlation with the ear temperature readings, (3) an imager temperature threshold should be
determined by the environmental factors, outdoor condition in particular, the physiological site offset and the performance characteristics of
thermal imager to warrant the most accurate and reliable screening operation. Conclusions: The analysis suggested that the thermal imager
used holds much promise for mass blind screening when the readings from a specific region have a good correlation with the ear temperature.
From the regression analysis, the best reading is taken from the maximum temperature in the eye region, followed by the maximum
temperature in the forehead region. With ROC analysis, a randomly selected individual from the fever group has a test value larger than that
for a randomly selected individual from the normal group in 97.2% of the time. The test can distinguish between the normal and febrile
groups and an optimum threshold temperature for the thermal imager can be found. The pre-set threshold cut-off temperature for the current
thermal imager was found to be 36.3jC with reference to the associated environmental condition. Any temperature readings that exceed this
reading will trigger off the alarm and a thermometer will be used to verify the whether the person is having fever.
D 2004 Elsevier Inc. All rights reserved.

Keywords: SARS; Avian flu; IR system; Bio-statistics; Fever; Threshold temperature

Introduction or oral temperatures so as to pick out those with fever and

send them for further clinical evaluation for SARS and
The cardinal symptoms of SARS and bird flu are fever hence curb community spread of the disease. Oral and aural
(de Jong et al., 1997; Ksiazek et al., 2003; Peris et al., 2003; temperature measurements are accurate but are fairly ‘‘in-
Rota et al., 2003) and this has led to temperature monitoring vasive’’, time-consuming, labour-intensive and skill-depen-
being practised at healthcare institutions, public areas and dent. The ideal device for fever screening should be speedy,
private establishments where crowds are expected. These non-invasive and be able to detect accurately those with
fever-screening stations employ personnel to take the aural fever with minimal inconvenience and disruption of human
traffic. As a first line of defense, infrared (IR) thermal
imaging has the potential to fulfil these functions and can
* Corresponding author. School of Mechanical and Production
Engineering, Nanyang Technological University, 50 Nanyang Avenue,
serve as a tool for mass screening for fever. However, there
Singapore 639798, Singapore. Fax: +65-6791-1859. is currently lack of scientific evidence to support this
E-mail address: [email protected] (E.Y.K. Ng). application. As quoted in the Canada National Post (Sep-

0026-2862/$ - see front matter D 2004 Elsevier Inc. All rights reserved.
doi:10.1016/j.mvr.2004.05.003
 E.Y.K. Ng et al. / Microvascular Research 68 (2004) 104–109 105

tember 24, 2003): ‘‘SARS scanners praised as placebo.

Health Canada report: Cost $2 million, of the 462,000
people screened in the first full month of operation, the
machines found 341 had fevers, but uncover no SARS
cases, yet ‘build confidence’’’. Current IR systems in use
at various boarder checkpoints have not been scientifically
validated particularly in regards to the false-negative rate.
As a result, they may create a false sense of security by
underestimating the number of febrile (and possibly
infected) individuals. The unadjusted mode threshold tem-
perature setting in a thermal imager needs to correct the
difference between the skin and core body temperatures. It
then has to take into account the effects of ambient con-
ditions and the thermal imager’s performance parameters. Fig. 2. A typical high fever (mean aural: 39.2jC) subject.
This paper studies two possible spots on the face (forehead
vs. inner corner region around eyes, other parts are either It detects infrared radiation emitted from the human skin
mostly covered or too inappropriate to scan) which yield IR surfaces and converts it into electrical signals, which are
skin temperatures that correlate/proxy with the core body then used to display a temperature profile/colour graphically
temperature (benchmark with averaged both ears tempera- and/or a temperature reading of a particular area of interest
ture, fever if z37.7jC in an adult using Braun Thermoscan (see Figs. 1 and 2). The subject is deemed to have an
IRT 3520+) for mass blind screening of fever to complement elevated temperature if the facial temperature is above the
and not to replace the conventional thermometers. threshold setting. Different factors are known to affect the
temperature reading. This may be from the thermal imager,
variation in the operating environment and the screened
Infrared camera technology subject. Thermal imagers have different degrees of temper-
ature drift between self-correction (Fig. 3), uniformity
IR imaging is a physiological test only rather than an within field of view, minimum detectable temperature
anatomical test such as X-rays or CT scan. The test is non- difference, error and stability of threshold temperature
invasive, and the camera and operator can be positioned at a (Fig. 4), distance effect and detector sizes.
distance from the subject to be screened (Ng and Sudharsan, The screened subject may contribute to the errors. False-
2001). The images show areas of both inflammations negative readings may be seen in subjects with heavy make-
(usually hot) and nerve dysfunction (usually cold) in the up or who are on medication. It is likely that hypothermia
patterns produced. As such it is possible to use infrared could be just as suspicious as hyperthermia. He/she who is
thermography for the detection of suspected SARS or bird perspiring heavily will present as significantly cooler due to
flu patients in a crowd as an early warning indicator due to evaporative cooling effect. False-positive results may be
its capability of detecting individuals with elevated (higher seen in subjects who are pregnant, menstruating, on hormone
than normal) temperature. replacement therapy, have recently consumed alcohol or hot
Usually, the thermal imager is ideally intended to operate drinks or have undergone physical exertion before screening.
in a stable indoor environment defined as an ambient Hence, it is useful to determine the cut-off temperature
temperature ranging from 20jC to 25jC and stability of setting, which is based on scientific and empirical result
F1jC, with a relative humidity ranging from 40% to 75%. obtained in this work.
Regression Analysis and Receiver Operating Characte-
ristics (ROC) will be used to analyse the data collected from

Fig. 3. An example of the drift of a typical thermal imager’s temperature

Fig. 1. A typical non-fever (mean aural: 37.0jC) case. reading between auto-adjustments (Standard Technical Reference, 2003).
106 E.Y.K. Ng et al. / Microvascular Research 68 (2004) 104–109

Fig. 4. An example of the overall system accuracy based on the various sources of uncertainty (Standard Technical Reference, 2003).

the thermal imager and conclusive results will be drawn Regression analysis
from them, the two very important pieces of information
from the study. Regression analysis is a statistical method for obtaining
the unique line that best fits a set of data points (Dupont,
 The specific region on the face (e.g., eye inner-corner 2002; Golberg and Cho, 2004). It fits the data by minimis-
region vs. forehead), which will provide a consistent ing the sum of the squares of the estimation errors. Each
correlation factor concerning the actual body temperature. error is the distance measured from the regression line to
 The optimal pre-set cut-off temperature for the thermal one of the data points. A regression analysis has two types
imager as the upper limit for normal healthy temperature of variables: the independent and the dependent variables.
(not to be confused with body or core temperature). The independent variable in our analysis is the core
Anyone whose skin surface temperature exceeds this temperature (measured using an ear clinical thermometer:
temperature will be suspected of having fever. When a Braun Thermoscan IRT 3520+) while the dependent vari-
fever like symptom is detected, the subject’s body able is the skin surface temperature (detected by the
temperature is taken with a thermometer and then checked thermal imager).
for respiratory distress symptoms. With the MedCalc software (ROC MedCalc (ver. 7),
2004), a regression equation (regression model, equation of
approximating curve) is selected. The linear least-square
Methodology equation Y = a + bX is chosen because it is the most
appropriate for the analysis (Fig. 5).
Data collection was carried out in the Emergency Depart- The R-squared is a very important coefficient to consider
ment, Tan Tock Seng Hospital (the designated SARS center in performing a Regression Analysis. It is a number between
in Singapore), the Singapore Civil Defense Forces and Civil 0 and 1, and is often described as a measure of the
Aviation Authority to study the relationship between the explanatory power of the regression; that is, the degree to
forehead and eye-vicinity temperatures and actual body which changes in the dependent variable can be predicted by
temperature. The total initial blind sample size collected change in the independent variable. These statistical data are
was 85 ‘febrile’ and 417 ‘normal’ cases. The scanner used
was the handheld radiometric IR ThermaCAM S60 FLIR
system (FLIR Systems, 2004). The focal length from subject
to scanner was 2 m and the duration of time patients must be
scanned was 3 s. The detector is a focal plane array, uncooled
microbolometer 320  240 pixels with a thermal sensitivity
of 0.08jC at 30jC, spectral range of 7.5 – 13 Am and
measurement accuracy at F2% of the real-time reading.
The mean temperature of a skin surface is measured from
the field of view of a thermal imager with an appropriate
adjustment for skin emissivity and it may vary from site to
site in the range of 0.94 – 0.99 (0.98 is used here). The skin
temperature is lower than the normal 37jC body temperature
because of well-studied heat evaporation, conduction and
convection principles. In fact, temperature measured on the
skin surface is a function of the temperature of an internal
organ, heat properties of tissues separating this organ from
the body surface and heat emissivity of the skin. Fig. 5. Temperature distribution of body vs. skin (eye range).
 E.Y.K. Ng et al. / Microvascular Research 68 (2004) 104–109 107

important because we would need to predict the core

temperature (independent variable) based on the tempera-
ture obtained from the thermal imager (dependent variable).
A more reliable regression is one that has an R-squared
close to 1 and vice-versa.

Receiver operating characteristics curve

ROC curve is a plot of true positive versus false-

positive results, usually in a trial of a diagnostic test
(Shapiro, 1999). It is a graphical means of assessing the
ability of a screening test to discriminate between healthy
and febrile persons. This will assist the clinical researcher
in accessing how accurate a particular laboratory test is in
identifying fever case. This is essential in analysing a test
with two populations, one with fever and one without,
since it is seldom in a case that there will be a perfect
and obvious separation between the two populations, that
Fig. 6. A typical ROC curve (eye range).
is, the distribution of the test results will overlap (Sha-
piro, 1999).
From Regression Analysis, it is known that the par-
ticular area on the skin surface that will produce the most
Results and discussion consistent results concerning the core temperature (aural)
is the area around the eye. More specifically, it is the
Regression data analysis maximum temperature in the eye region. This is found
from the coefficient of determination, which is 0.5509.
Notice that for those who are wearing contact lenses,
Eye range (max) their eyeballs with lenses would not affect the tempera-
Dependent Y: Eye Range (Max) ture profiles on screening the temperature at the inner
Independent X: Ear Temp corner of the eyes. In fact, the eyeballs are cooler than
Sample size = 310
other parts of the body based on medical understanding.
Coefficient of determination = 0.5509
Residual standard deviation = 0.7061 It is also shown that the minimum temperature at the eye
- REGRESSION EQUATION - region will not yield such a high level of consistency
Y = 4.1972 + 0.8509 X where the coefficient of determination is only 0.0622. A
similar analysis was done on the forehead region. The
Eye range (min)
maximum temperature spot yields a coefficient of deter-
Dependent Y : Eye Range (Min)
Independent X: Ear Temp mination of 0.4974 while the minimum temperature spot
Sample size = 310 yields 0.1170. Thus, the eye range region allows a better
Coefficient of determination = 0.0672 site to correlate to (or monitor) the body temperature. In
Residual standard deviation = 1.7788 fact, there are quite a few arteries around the eye (the
- REGRESSION EQUATION -
ophthalmic artery is in vicinity to the lacrimal caruncle
Y = 11.7507 + 0.5194 X
and it is connected to the optic nerve) (Virtual hospital,
Forehead (max) 2004). Thus, a small area of skin near the eyes and noise
Dependent Y: Forehead (Max) offers the body’s core temperature to be measured since
Independent X: Ear Temp the thin skin in this area has the highest amount of light
Sample size = 310
energy making it a preferred point of entry for the brain
Coefficient of determination = 0.4974
Residual standard deviation = 0.8108 temperature tunnel.
- REGRESSION EQUATION -
Y = 2.7484 + 0.8776 X ROC data analysis
Forehead (min)
Dependent Y: Forehead (Min)
Eye range (max)
Independent X: Ear Temp
Sample size = 310 Fig. 6 illustrates the true positive (fever based on eye-
Coefficient of determination = 0.1170 range region) rate in function of the false-positive rate at
Residual standard deviation = 1.6698 different cut-off temperature points. This nonparametric
- REGRESSION EQUATION - approach is free of distributional assumptions in that it
Y = 5.7581 + 0.6612 X
depends only on the ranks of the observations in the
108 E.Y.K. Ng et al. / Microvascular Research 68 (2004) 104–109

combined sample, but the resulting empirical ROC curve specificity from the ROC analysis results. In other words,
is a series of horizontal and vertical steps (in the absence the chosen threshold temperature must have high values of
of ties), which may be quite jagged as can be seen in sensitivity and specificity. For the maximum eye region
Fig. 6. The area under the ROC curve (AUC) with temperature, to detect an adult aural temperature of 37.7jC
standard error and 95% CI is 0.972. This means that a and above, a threshold temperature of 36.3jC is selected
randomly selected individual from the positive (fever) and this allows sensitivity of 85.4% and specificity of 95%.
group has a test value larger than that for a randomly This temperature is suggested for the current system and
selected individual from the negative (normal) group in environmental condition as no other temperatures will be
97.2% of the time. Therefore, the current test does have able to yield such high values. For an example, if 36.7jC
an ability to distinguish between the normal and febrile was chosen then sensitivity would be 72.9% with specificity
groups. of 99.6%. ROC analysis was also performed on the maxi-

VARIABLE = Eye Range Max

CLASSIFICATION VARIABLE
Diagnosis
POSITIVE GROUP
Diagnosis = 1
Sample size = 48
NEGATIVE GROUP
Diagnosis = 0
Sample size = 262
Area under the ROC curve = 0.972
Standard error = 0.017
95% Confidence interval = 0.947 to 0.987
Criterion Sens. (95% C.I.) Spec. (95% C.I.) +LR LR
> =33 100.0 (92.5 – 100.0) 0.0 (0.0 – 1.4) 1.00
> 33 100.0 (92.5 – 100.0) 0.4 (0.1 – 2.1) 1.00 0.00
> 36.1 85.4 (72.2 – 93.9) 92.7 (88.9 – 95.6) 11.78 0.16
> 36.2 85.4 (72.2 – 93.9) 93.9 (90.3 – 96.5) 13.99 0.16
> 36.3* 85.4 (72.2 – 93.9) 95.0 (91.7 – 97.3) 17.21 0.15
> 36.4 83.3 (69.8 – 92.5) 96.2 (93.1 – 98.2) 21.83 0.17
> 36.6 75.0 (60.4 – 86.3) 98.5 (96.1 – 99.6) 49.12 0.25
> 36.7 72.9 (58.2 – 84.7) 99.6 (97.9 – 99.9) 191.04 0.27
> 37 66.7 (51.6 – 79.6) 99.6 (97.9 – 99.9) 174.67 0.33
Sens. = Sensitivity, Spec. = Specificity
+LR = Positive likelihood ratio, LR = Negative likelihood ratio

Forehead range (max)

VARIABLE = Forehead Max
CLASSIFICATION VARIABLE
Diagnosis
POSITIVE GROUP
Diagnosis = 1
Sample size = 48
NEGATIVE GROUP
Diagnosis = 0
Sample size = 262
Area under the ROC curve = 0.960
Standard error = 0.020
95% Confidence interval = 0.932 to 0.979
Criterion Sens. (95% C.I.) Spec. (95% C.I.) +LR LR
> =32.3 100.0 (92.5 – 100.0) 0.0 (0.0 – 1.4) 1.00
> 35.7 91.7 (80.0 – 97.6) 92.0 (88.0 – 95.0) 11.44 0.09
> 35.8* 89.6 (77.3 – 96.5) 94.3 (90.7 – 96.8) 15.65 0.11
> 35.9 83.3 (69.8 – 92.5) 95.4 (92.1 – 97.6) 18.19 0.17
> 36 81.2 (67.4 – 91.0) 97.3 (94.6 – 98.9) 30.41 0.19
> 36.1 75.0 (60.4 – 86.3) 98.5 (96.1 – 99.6) 49.12 0.25
> 36.3 70.8 (55.9 – 83.0) 99.6 (97.9 – 99.9) 185.58 0.29

With ROC curves analysis, it is desired that an optimum mum forehead region temperature. A temperature of 35.8jC
threshold temperature for the thermal imager can be found. is thus recommended for the same reason. Finally, the
This setting is dependent on the values of sensitivity and thermal scanner temperature threshold should be determined
 E.Y.K. Ng et al. / Microvascular Research 68 (2004) 104–109 109

by the environmental factors, the physiological site offset Acknowledgments

and the performance characteristics of the thermal imager to
warrant the most accurate and reliable screening operation. The first author would like to express his appreciation to
Study shows that the human skin surface temperature members of the Ad hoc Technical Reference Committee on
correlates to the body core temperature to certain extend. Thermal Imagers under Medical Technology Standards
However, it is risky to use a fixed physiological site offset to Division by SPRING, and Ministry of Health, TTSH of
correlate both temperatures for the threshold temperature National Health Group, Singapore for sharing of their views
setting as the skin surface temperature changes at a different and interests on ‘‘Thermal Imagers for Fever Screening-
time of the day and at a different environment. Selection, Usage and Testing’’.

Conclusion References

The analysis suggested that the thermal imager used de Jong, J.C., Claas, E.C., Osterhaus, A.D., et al., 1997. A pandemic warn-
here holds much promise for mass blind screening when ing? Nature 359, 544.
Dupont, D.W., 2002. Statistical Modelling for Biomedical Researchers: A
the readings from a specific region have a good correlation Simple Introduction to the Analysis of Complex Data. Cambridge Univ.
with the ear temperature. From the Regression Analysis, Press, UK (Chps. 2 – 3).
the best reading is taken from the maximum temperature in FLIR Systems: http://www.flir.com (accessed 4th May 2004).
the eye region, followed by the maximum temperature in Golberg, M.A., Cho, H.A., 2004. Introduction to Regression Analysis. WIT
the forehead region. In addition, the pre-set threshold cut- Press, Southhampton, England.
Ksiazek, T.G., Erdman, D., Goldsmith, C.S., Zaki, S.R., Peret, T., Emery,
off temperature for the current thermal imager has been S., Tong, S.Z., et al., 2003. A novel coronavirus associated with severe
found to be 36.3jC with reference to the associated acute respiratory syndrome. N. Engl. J. Med. 348, 1953 – 1966.
environmental condition. Any temperature readings that Ng, E.Y.-K., Sudharsan, N.M., 2001. Numerical computation as a tool to aid
exceed this reading will trigger off the alarm and a thermographic interpretation. Int. J. Med. Eng. Technol. 25 (2), 53 – 60.
thermometer will be used to verify the whether the person Ng, E.Y.-K., Kaw, G., Ng, K., 2004. Infrared thermography in identifi-
cation of human elevated temperature with Biostatistical and ROC
is having fever (Ng et al., 2004). analysis. The International Society for Optical Engineering (SPIE).
The WHO has repeatedly warns that SARS would ThermoSense XXVI-Symposium on Defense and Security, 12-16 April,
return this winter and the world has to be on the alert Orlando, FL, USA. SPIE, Bellingham, WA, pp. 88 – 97.
for the resurgence of the virus. Four full-blown ‘reacti- Peris, J., Lai, S., Poon, L., et al., 2003. Coronavirus as a possible cause of
vated’ SARS cases in Guangdong, China, were confirmed severe acute respiratory syndrome. Lancet 361, 1319 – 1325.
ROC MedCalc (ver. 7): http://www.medcalc.be (accessed 4th May 2004).
(winter) and fresh reports of bird flu crisis elsewhere in the Rota, P.A., Oberste, M.S., Monroe, S.S., et al., 2003. Characterization of a
region including Vietnam, South Korea, Taiwan, Japan, novel coronavirus associated with severe acute respiratory syndrome,
Indonesia, Thailand, Pakistan, Bangladesh, Laos, Guangxi Sciencexpress http://www.sciencemag.org/cgi/rapidpdf/1085952v1.pdf
and Cambodia. The timing could not have been worse, (accessed 4th May 2004).
with the expected mass migrating for the seasonal birds Shapiro, D.E., 1999. The interpretation of diagnostic tests. Stat. Methods
Med. Res. 8, 113 – 134.
from South back to North hemisphere. To prepare for Standard Technical Reference for ‘‘Thermal Imagers for Human Tempera-
future SARS or avian flu epidemics, reliable screening ture Screening Part 2: User’s implementation guidelines, July, 2004.
methods to detect febrile (and presumably infected) indi- TR 15 – 2, Spring, Singapore.
viduals need to be rapidly implemented at airports and Standard Technical Reference for ‘‘Thermal Imagers for Human Tempera-
immigration centers. An effective IR system (Standard ture Screening Part 2: Users’ implementation guidelines’’, TR 15-1,
Spring Singapore, (2004) (in press).
Technical Reference, 2003) with a strict operating protocol Virtual hospital, Atlas of Human Anatomy, Translated by: Ronald A. Berg-
(Standard Technical Reference, in press) should be adopted man and Adel K. Afifi: http://www.vh.org/adult/provider/anatomy/
to facilitate this screening process. atlasofanatomy/plate17/index.html (accessed 4th May 2004).