Journal of Rehabilitation Research and
Development Vol. 38 No. 1, January/February 2001
Pages 13–22
Testing the validity of erythema detection algorithms
Brian Riordan, MEBME; Stephen Sprigle, PhD, PT; Maureen Linden, MSBME
Center for Rehabilitation Technology, Helen Hayes Hospital, West Haverstraw, NY 10993
INTRODUCTION
Abstract—Dermatology has quantified skin color for monitoring progress of treatments. The most common and effective
means of erythema detection is visual inspection of the skin.
However, for people with darkly pigmented skin, erythema can
be masked by melanin. Tissue Reflectance Spectroscopy (TRS)
is a noninvasive method of quantifying skin color. Most commonly, TRS quantifies erythema caused by cosmetics, topical
ointments, UV light, or other irritants. Recently, TRS has been
used to characterize the presence of erythema due to reactive
hyperemia or Stage I pressure ulcers. The objective of this
study was to compare the reliability and validity of erythema
detection algorithms by determining their sensitivity and specificity. Two algorithms, Diffey and Helen Hayes Hospital
(HHH), had sensitivity exceeding 85% and specificity exceeding 75%, but most algorithms demonstrated adequate validity
across all subjects. The validity of the HHH algorithm did not
change with the skin pigmentation of the subject. The results of
this comparison will be useful to researchers interested in using
TRS to detect erythema in people with different skin pigment
levels.
The field of dermatology has long been interested in
inspecting skin color and changes in skin color during
and after treatment. One common noninvasive measurement is via tissue reflectance spectroscopy (TRS). The
TRS uses the characteristic absorption of light by the constituents of the skin to measure the amount of various
constituents present. A white light is shone on the skin
while detectors measure the returning light. From the
spectral characteristics of the light returned, the relative
amounts of each skin chromophore can be determined.
Most erythema studies using TRS target the effects that
cosmetics, topical ointments, UV light, or other irritants
on the skin (1,2). Recently, TRS has been used to characterize the presence of erythema due to reactive hyperemia
or Stage I pressure ulcers (3).
To detect and quantify the erythema, several algorithms using TRS absorption data have been reported in
the literature. These algorithms have been tested primarily with lightly pigmented subjects; hence, their ability to
detect erythema in darkly pigmented tissue remains
unknown. The objective of this study was to determine
the reliability and validity of erythema detection algorithms for all pigment levels. Five methods reported in
the literature, and a sixth, developed during ongoing
research with diabetic amputees, were investigated.
Validity was determined using the sensitivity and specificity of each. In this case, sensitivity is the ability to cor-
Key words: algorithm, erythema, melanin, skin pigmentation,
tissue reflectance spectroscopy.
This material is based on work supported by the National Institute on
Disability and Rehabilitation Research, Award #H133G50018; the
National Institutes of Health, Award #HD 31428-03; and the New York
State Department of Health.
Address all correspondence and requests for reprints to: Stephen Sprigle, PhD,
PT, Center for Rehab Technology, Helen Hayes Hospital, Rt. 9W, West
Haverstraw, NY 10993; email: gogators@compuserve.com.
13
14
Journal of Rehabilitation Research and Development Vol. 38 No. 1 2001
rectly identify tissue with erythema, and specificity is the
ability to correctly identify tissue without erythema.
The results of these comparisons will be useful to
researchers interested in using TRS to detect erythema in
people with different skin pigment levels. One specific
benefit of a robust erythema detection algorithm is the
development of an instrument for use by health care professionals to detect erythema. This can be useful in monitoring reactive hyperemia or detecting Stage I pressure
ulcers in deeply pigmented subjects. Detection of a Stage
I ulcer will allow timely intervention to prevent progression of the ulcer.
Tissue Reflectance Theory
The TRS uses the characteristics of light reflected by
the constituents of the skin to measure the amounts of
various constituents present. A white light is shone on the
skin while detectors measure returning light. The light is
divided into spectral components by a monochromator
and detected by a photomultiplier tube. To eliminate
transmission effects of the instrumentation, reflectance
from the skin is compared to reflectance from a white
standard. The result is a unitless quantity termed relative
reflectance. Absorption, or Log Inverse Reflectance
(LIR), refers to the amount of light not returned from the
skin. This is calculated as the name suggests, with the
result expressed in absorption units (au):
Absorption51/log10(relative reflectance)
[1]
The theory of TRS is based upon a simple anatomical
model (4). Light passes through the epidermis (melanin
layer) and a plexus of blood vessels in the dermis (hemoglobin (Hb) layer) before being reflected off collagen in
the lower dermis. This model, shown in Figure 1, highlights the human tissue chromophores within the visible
spectrum that influence the measurement of erythema:
Hb, oxyhemoglobin (oxyHb), and melanin.
The Hb absorbs light with a characteristic curve showing broad bands of absorption in the green portion of the
spectrum. The oxyHb has two absorption maxima at 542 nm
and 574 nm wavelengths. De-oxyhemoglobin (deoxyHb)
shows a single maximum at 545 nm. Thus, TRS can theoretically produce information about both the amount of Hb
present and its degree of oxygenation. Melanin has a linearly decreasing curve in the spectral range from 500–700 nm.
The slope of this curve increases as the melanin content of
Figure 1.
a) Schematic structure of the skin; b) Optical skin model of three-layered structure with an outer melanin layer, an upper hemoglobin layer,
and a backing representing chromophore-free dermis. Reprinted with
permission from (Takiwaki S, Serup H. Measurement of erythema and
melanin indices. In: Serup H, Jemec GBE (editors). Handbook of noninvasive methods and the skin. Boca Raton: CRC Press; 1955. p. 378,
Figure 1 with caption), Copyright CRC Press Boca Raton, Florida.
an individual’s skin increases. For people with darkly pigmented skin, the absorption by melanin can be much greater
than that of Hb (Figure 2). This masking by melanin makes
it difficult to detect erythema.
Figures 3 and 4 illustrate the difference in TRS
spectra obtained from lightly and darkly pigmented individuals. The absorption curve of the individual with deep
pigmentation has greater overall amplitude (increased
light absorption) and the Hb “double hump” is not as easily discerned.
15
RIORDAN et al. Testing Validity of Erythema Detection Algorithms
Algorithms Tested
The use of TRS to detect pressure ulcers in people
with darkly pigmented skin is dependent on two processes:
1. a means to adjust or correct the signal for melanin
Ideally, melanin compensation would be independent of
the amount of blood present and effective for all levels of
pigmentation.
2. a robust erythema detection algorithm
Figure 2.
Characteristic curves of human skin chromophores.
Five methods of erythema detection were identified
in the literature. A sixth was developed during ongoing
research on diabetic amputees and nondisabled controls.
Each of these algorithms uses its own method of calculating an index of erythema. Most, but not all, also quantify and correct for melanin. These methods are described
briefly below. For detailed descriptions of these algorithms, the reader should consult the cited references.
1. Dawson
Dawson based melanin compensation on the difference of absorption from 650 nm to 700 nm. He averaged
absorption values at 645, 650, and 655 nm and subtracted the average of measurements at 695, 700, and 705 nm
(5). Because of differences in the response of our instrument, we calculated the index based on the difference
between an average of ten wavelengths from 640–650 nm
and ten from 690–700 nm, respectively.
Figure 3.
Absorption spectral response curve for an individual with lightly pigmented skin.
MDaw5(L640-650-L690-700)*100
[2]
where Lx2y is the average absorption in the wavelength
range from x to y nm.
Dawson’s erythema index, EDaw, is a parameter that
is proportional to the area under the Hb absorption curve
when an artificial baseline is drawn between 510 nm and
610 nm.
[
]
EDaw5100 A56011.5*(A5401A575)22(A5101A610)
Figure 4.
Absorption spectral response curve for an individual with deeply pigmented skin.
[3]
where An is the absorption of the spectrum at wavelength
n (5).
E is corrected for melanin by adding the melanin index
scaled by a factor g(=0.04). This factor was empirically
determined based on the assumption that lightly and darkly
pigmented subjects have similar blood content.
16
Journal of Rehabilitation Research and Development Vol. 38 No. 1 2001
EC5Edaw(12gMDaw)
[4]
2. Ferguson-Pell
Ferguson-Pell defined a melanin index (IMEL) as
the slope of the regression line for the bloodless
(blanched) absorption spectrum from 500–600 nm (6).
After the slope of the bloodless spectrum is subtracted
from every spectral curve, the response is attenuated by a
multiplication factor, which was a function of IMEL and
was empirically determined based on the assumption that
lightly and darkly pigmented subjects had similar skin
blood content (6).
Ferguson-Pell than quantified erythema by the
amount of Hb in the skin surface (IHB) as developed by
Feather (7). The IHB is based on the isobestic wavelengths of Hb, or the wavelengths where Hb absorption is
not affected by its oxygenation level (absorption of
oxyHb=absorption of deoxyHb at these wavelengths).
The formula for IHB is shown below (Equation 5), where
An is the absorption of the spectrum at wavelength n after
attenuation compensation for melanin, as described
above (7).
[
IHB550*
A5452A522 A5682A545
2
23
23
]
[7]
“True” IHB equals the corrected IHB plus the IMEL
expressed in ug/cm2, scaled by an empirically determined
factor (8).
IHBt5IHBc1.00047*MHaj
[8]
4. Diffey
Diffey noted that Hb had high absorption in the
green spectrum and low absorption in the red spectrum,
and, further, that erythema caused by vasodilation produced significant increases in green absorption and little
change in red absorption. His premise was that differences between red and green absorption were due solely
to Hb content and that melanin compensation would be
accomplished by comparing the two spectral regions (9).
(
Edif 5 log10
)
REF(635nm)
REF(565nm)
[9]
where REF(x)=Reflectance at wavelength x nm.
[5]
3. Hajizadeh-Saffar
Hajizadeh-Saffar also used the work of Feather to
calculate IHB, but first corrected for light backscatter in
the epidermis. Melanin compensation is based on the
slope of the absorption curve from 650–700 nm. Initial
estimates are used to compensate for blood content and
oxygenation in the skin (8).
M6755{(A7002A650)/501
(0.060[1-SaO2c/100]10.010)IHBc/80}*100
MHaj5(0.0542M675)*120.1 mg/cm2
[6]
where: An is the corrected absorption at wavelength n;
SaO2c is an index of blood oxygenation corrected for
backscatter; and, IHBc is the Hb index corrected for
backscatter.
The melanin index can be expressed in terms of the
concentration of synthetic melanin required to produce
the same response in vitro.
5. Tronnier
Tronnier’s erythema index is based on the difference
between red and green reflectance at a control site and an
erythematic site. Melanin compensation is achieved by
comparing two sites (10):
ETro 5 (G 2 R) 2 (GO 2 RO)
[10]
where: G, Go=reflectance at 545 nm (erythematic, control
site) R, Ro=reflectance at 661 nm (erythematic, control site)
One should note that Tronnier subtracted reflectance
values at two sites, the first known to have erythema caused
by ultraviolet radiation while the other a control site. The
control site was measured primarily to evaluate the pigmentation of a subject. This technique produces a single
value, characterizing erythema of one site relative to the
other.
Because our goal in this analysis is to develop an
algorithm useful for determining whether a site has erythema, strict application of Tronnier’s algorithm is not
17
RIORDAN et al. Testing Validity of Erythema Detection Algorithms
appropriate. Instead, we subtracted green reflection from
red reflection using the formula below and quantified the
differences between sites with percent differences and Zscore calculations.
where: G=reflectance at 545 nm; R=reflectance at 661
nm.
ETro 5 (G 2 R)
[11]
6. Helen Hayes Hospital (HHH)
A concentration-independent curve for melanin was
calculated from in vitro data. This slope is scaled for each
individual by the difference in absorption from 500 nm to
625 nm. The strengths of this approach include its focus
over the area of interest and its ability to accommodate
both lightly and darkly pigmented skin. Its weaknesses
include an incomplete understanding of how melanin
concentration changes over adjacent areas of tissue (i.e.,
a control site and an adjacent “red” site).
The melanin curve is subtracted from the spectrum.
The compensated spectrum is regressed using a standard,
concentration-independent absorption curve of Hb (in
vitro Hb response) as the regressor. The coefficient (beta
code this to match the equation and make a note on the
hard copy) of this regression is a measure of Hb concentration. If necessary, this model can be expanded to
include separate regressors corresponding to oxyHb and
deoxyHb.
METHODS
Algorithms were tested by measuring tissue
reflectance at an erythematic site and at two adjacent control sites. As mentioned previously, the use of TRS to
measure erythema depends upon determining melanin
and Hb, so the protocol was designed to permit calculations of both. Each detection algorithm was applied to the
same reflectance data. In addition to the six algorithms
defined above, a seventh algorithm was tested that
applied the HHH melanin correction approach to erythema detection with the Dawson algorithm.
Equipment
A Monolight 6800 series spectrophotometer (Rees
Instruments, Smyrna, GA), consisting of scanning monochromator, photodetector, and light source with a fiber
optic cable, was used for TRS data collection. A rotating
diffraction grating dispersed incident light into component wavelengths that were then converted into electrical
signals by the 6117 photodetector. The grating blaze
wavelength, 500 nm, yields an optimal operating range
from 350–1100 nm, a dispersion of 10 nm/mm, and a resolution of 1 nm in the visible spectrum. A 6162 stabilized
tungsten halogen lamp provided the light source. A second order blocking filter (WG345, range 350–640 nm)
negated the aliasing effects produced by the rotation of
diffraction grating.
The bifurcated fiber optic cable (Fiber Guide
Industries, Caldwell, ID) consisted of 2,370 borosilicate
fibers. Roughly half the fibers transmitted the incident
light, while the remaining fibers carried light back to the
photodetector. The end of the fiber optic cable was fitted
with a probe with a 3-mm diameter optic aperture. The
aperture side of the probe had a slightly convex shape to
minimize edge effects during pressure application, and a
merosphere was affixed to the top of the probe to interface with the pressure application system.
Localized erythema was induced on the shank of
subjects using a pneumatic indentor that applied 150
mmHg. The pressure application system consisted of a
computer-controlled pneumatically driven piston. The
piston is equipped with a load cell (Entran ELF-C100010), which provided feedback to the controller and maintained 150-mmHg pressure while accommodating for
slight body movements during monitoring. LabTech
Notebook Pro, Ver 10 (Laboratory Technologies Corp,
Wilmington, MA) was used to control the system. Details
of the indentor system have been previously published in
this journal (11).
Data Collection
A convenience sample of 20 subjects was recruited
from hospital staff. The skin of each subject was classified according to a Munsell color chart (5YR; reference
12). The Munsell chart consists of 33 tiles of color that
are scaled on value (2.5–8) and chroma (1–8). The value
quantifies the darkness of the skin, with lower numbers
indicating darker skin, while the hue quantifies the ruddiness of complexion, with higher numbers indicating reddish complexion. Of the 20 subjects tested, 9 were
Caucasian (values from 7–8), 3 were Hispanic (values
from 6–7), and 8 were Black (values from 3–6).
A consent form that described the project and risks
was provided to each subject. After informed consent was
obtained, subjects were positioned in a semi-recumbent
position on a mat. The shank of the test leg was exposed
18
Journal of Rehabilitation Research and Development Vol. 38 No. 1 2001
to an area near the tibial flare. This area was tested
because it is prone to reactive hyperemia and pressure
ulcers in patients with transtibial amputation who ambulate with patellar tendon- bearing prostheses. Test sites
that were free of hyper- or hypopigmentation and scarring
were identified, cleaned, and shaved, if necessary. Three
electrode adhesive rings were affixed to the skin, one
marking the test area and the other two at the two adjacent control sites. The electrode adhesive rings were
slightly larger than the spectrometer aperture (3-mm
diameter), allowing repeated measures to be performed at
each test site. The TRS measurements were taken at each
control site before inducing erythema, to permit calculation of melanin.
Localized erythema was induced on the shank of
subjects using a pneumatic indentor that applied a 150mmHg pressure for 3 min. Reflectance data from each
site was measured with the spectrometer for 30 s at low
contact force and 10 s with a contact force sufficient to
blanch the skin. The optic head of the spectrophotometer
was removed and re-seated on the skin between each
measure, to create independent measures at each site.
The protocol was based upon the potential for using
TRS as a means to identify sites at risk for pressure ulcer
development or any other clinical monitoring. Therefore,
TRS measurements were done for short periods of time
(40 s) and a researcher positioned and held the optic head
against the skin, using a hand-held pressure application
system. This pressure application system allowed positioning of the optical head such that light was emitted
normal to the skin surface, while ensuring that the pressure applied was less than 40 mmHg. Previous investigation had concluded that, for this optical head and this
anatomical location, 40 mmHg does not cause blanching
of the skin. The pressure application system also ensured
a pressure greater than 150 mmHg during the blanching
of the skin required for calculation of IMEL. The minor
variations caused by hand-held monitoring assisted in
testing the robust nature of each algorithm.
Repeatability was tested by measuring each site
twice, in the following sequence:
Control site 1
Control site 1
Control site 2
Control site 2
Erythema site
Erythema site
The control site measures were performed while the
test site was being loaded. Because reactive hyperemia is
a transient event, the subsequent measures of erythema
were not expected to yield the same Hb content; rather
they were used to test the detection algorithm’s ability to
identify the presence of erythema compared to control
sites. Reactive hyperemia typically lasts between 50–70
percent as long as the ischemic event (13), in this case 90
s, so the second measure of erythema at the test site falls
well within this period.
Data Analysis
This data collection permitted sensitivity to be determined by comparing the erythema test site to the control
sites, and allowed specificity to be determined by comparing the control sites. In other words, sensitivity determined whether the test sites were different than the
control sites (an erythema site was determined to have
erythema or a true positive result) and specificity determined whether control sites were not different (a non-erythema site was determined to be non-erythematic or a true
negative response). The melanin correction and erythema
detection algorithms were performed as described above.
Within each detection algorithm, four detection criteria were studied. Three Z-scores and a 3-percent change
in Hb content were used within each algorithm to determine the optimal criteria.
Z scores were calculated using the formula:
Z=(msiteA2msiteB)/ssiteB.
Percent
change
was
calculated
using:
D=(msiteA2msiteB)/msiteB.
Z scores have the benefit of using both the mean and
standard deviations and represent the normalized value
from a distribution. Percent change uses only the mean at
both sites, but is more intuitive for most persons.
Within each detection algorithm and detection criteria, the data were categorized into a 2 3 2 logic table
divided into True-positive, True-negative, False-positive,
and False-negative designations. Tabulated results were
used to calculate:
Sensitivity=True-positive/(True-positive1False-negative)
Specificity=True-negative/(True-negative1False-positive)
where: True-positive=erythema sites identified as erythema
sites
True-negative=non-erythema sites identified as non-erythema sites
False-positive=non-erythema sites identified as erythema
False-negative=erythema sites identified as non-erythema
Reliability was determined by analyzing the two
independent measures at each control site. Because neither site was erythematic, the repeated measures should
19
RIORDAN et al. Testing Validity of Erythema Detection Algorithms
be consistent quantitatively and should not differ by the
threshold criteria. In other words, a site should not be
deemed different from itself after repeated measures.
Therefore, repeatability was calculated in 2 manners: 1)
Intraclass correlation coefficient (ICC) to measure the
quantitative consistency of the erythema measurement
and specificity of the control sites [true-negative/(falsepositive1true-negative)] to determine how often repeated measures of the same site differed by the threshold
(false-positive result).
RESULTS
The optimal Z-score and percent difference for each
of the algorithms and the resulting validity measures are
contained in Table 1. The “optimal” criteria were determined using four conditions: 1) sensitivityt0.80 for the
three groups (All Subjects, Dark Skin Subjects, and Light
Skin Subjects); 2) sum of sensitivity1specificity value;
3) sensitivitytspecificity; and, 4) dark skin total sum.
Appendix A lists the sensitivity and specificity of the six
criteria used for each algorithm. Table 2 lists the reliability results, including ICC and the specificity of the control sites for each optimal criterion.
The HHH was the only algorithm to exceed 80 percent
sensitivity and specificity levels for both the Z-score and percent difference thresholds. The Diffey algorithm, using Z=2.5
threshold, approached these sensitivity and specificity levels
with values at 88 percent and 78 percent, respectively.
The reliability of all algorithms was high if using
ICC as the measure, but only Dawson, IHB, and HHH
Table 1.
Algorithm validity
Algorithm
Z Score
Criteria
Dawson
Dawson-HHH
Diffey
Hajizadeh
HHH
IHB
Tronnier
Algorithm
3
2
2.5
1
4
1
3
%
Difference
Criteria
Dawson
Dawson-HHH
Diffey
Hajizadeh
HHH
IHB
Tronnier
10
5
20
20
40
20
20
All Subjects
Sensitivity
Specificity
0.809
0.827
0.88
0.691
0.9
0.609
0.818
0.765
0.758
0.784
0.883
0.846
0.901
0.79
All Subjects
Sensitivity
Specificity
0.818
0.836
0.845
0.832
0.873
0.84
0.845
0.769
0.747
0.741
0.68
0.887
0.703
0.698
Dark-Skin Subjects
Sensitivity
Specificity
0.815
0.815
0.852
0.667
0.944
0.611
0.796
0.718
0.667
0.744
0.897
0.859
0.91
0.718
Dark-Skin Subjects
Sensitivity
Specificity
0.741
0.815
0.796
0.87
0.907
0.842
0.796
0.782
0.744
0.769
0.636
0.91
0.704
0.705
Light-Skin Subjects
Sensitivity
Specificity
0.804
0.839
0.907
0.714
0.857
0.607
0.839
Light-Skin Subjects
Sensitivity
Specificity
0.893
0.857
0.893
0.8
0.833
0.839
0.893
Table 2.
Algorithm reliability
Algorithm
ICC
Reliability
Using Z Scores
Dawson
Dawson-HHH
Diffey
Hajizadeh
HHH
IHB
Tronnier
.795
.838
.891
.655
.812
.998
.921
.63
.53
.65
.37
.75
.72
.60
0.81
0.845
0.821
0.869
0.833
0.893
0.857
Using % Difference
.75
.52
.63
1.0
.88
.43
.55
0.81
0.75
0.714
0.714
0.861
0.702
0.69
20
Journal of Rehabilitation Research and Development Vol. 38 No. 1 2001
demonstrated reliability exceeding 70 percent when calculated with respect to the optimal threshold. This difference can be attributed to the meanings of each respective
reliability measure. The ICC reflects the variance of values within each repeated measure compared to the variance across subjects. Conversely, repeatability with
respect to difference thresholds is obviously dependent
on the threshold magnitudes and how they relate to the
variation of same-site, repeated measures.
DISCUSSION AND CONCLUSIONS
The results indicate that certain algorithms are able
to detect erythema in skin of varying pigmentation. To
date, few erythema research studies addressed skin pigmentation. This conclusion may be used in the development of clinical tools to detect erythema.
Often, sensitivity and specificity are inversely related in clinical tests that measure a continuous variable
(e.g., blood sugar, blood pressure). Since erythema is a
continuous variable, changing the threshold criteria used
to indicate erythema can alter the sensitivity and specificity. This project tested six thresholds and designated
“optimal” thresholds for each algorithm according to the
criteria listed above. Sensitivity was deemed more important than specificity but the sum total of both were also
included.
Calculating both sensitivity and specificity is important in the study of detection algorithms because these
parameters affect Type I and Type II errors. Balancing
these errors depends on many factors such as costs of the
test and costs associated with false positives. In certain
instances, one may need to maximize sensitivity in order
to ensure that sites with erythema are identified with minimal error. Alternatively, maximizing specificity would
ensure that skin free of erythema is identified accurately.
If a detection algorithm was used clinically to identify
tissue at risk for ulcer formation, the risk of incorrectly
identifying erythema when it was not present (false-positive) would probably lead to increased monitoring of that
site. While this error might lead to additional staff time, it
should not have any adverse effects on the patient.
Conversely, if an erythema site was incorrectly identified as
not erythematic (false-negative), the patient is at risk.
Therefore, in judging algorithms, sensitivity was deemed
more important than specificity because the higher the sensitivity the lower the risk of a false-negative finding, and
thus a lower risk to the patient.
Two algorithms, HHH and Diffey, exceeded 85 percent sensitivity and the other 5 algorithms had criteria
that led to a sensitivity exceeding 80 percent. These are
acceptable values for clinical tests, especially given the
concomitant specificity values of most of the algorithms
(>75 percent).
The Z=1 criteria of IHB and Hajizadeh are suspect
as viable thresholds, as are the 10 percent and 5 percent
difference criteria for the two Dawson algorithms. The
inherent variability of TRS, as with many physiological
variables, leads to questioning the use of very low detection thresholds. In other words, a low detection threshold
might lack robustness in a clinical setting as opposed to a
controlled research setting. This intuition, however, was
not completely corroborated by the reliability measures.
While the Hajizadeh algorithm returned relatively high
false-positive results using a Z=1 threshold, IHB exhibited a specificity of 72 percent. Reliability of the DawsonHHH algorithm (52 percent) was affected by the low
percent-difference threshold, but the Dawson algorithm
exhibited an acceptable reliability (75 percent).
The limitations of the study include both risks to
external and internal validity. In this study, erythema was
induced via a reactive hyperemic response to localized
ischemia. While this is consistent with pressure ulcer formation, these results might not generalize to erythema
induced by other means (e.g., topical ointments, UV, full
limb ischemia). Similarly, the small number of subjects
might also limit external validity, although this study
used a variety of skin pigmentations to test the validity of
each algorithm. An internal validity risk might be due to
instrumentation bias. The influence on these algorithms is
unclear. Theoretically, the algorithms should act independently of the spectrometer used, as long as its accuracy,
resolution, and precision are adequate. The spectrometer
used in this study had adequate spectral range and a 1-nm
resolution.
21
RIORDAN et al. Testing Validity of Erythema Detection Algorithms
APPENDIX
Raw data
Criteria
Algorithm
Results
Z Score
Overall
Dark
Light
Sens
Spec
Sens
Spec
Sens
Spec
HHH
5
4
3.5
0.855
0.900
0.918
0.870
0.846
0.778
0.907
0.944
0.963
0.897
0.859
0.756
0.804
0.857
0.875
0.845
0.833
0.798
IHB
2
1.5
1
0.373
0.500
0.609
0.988
0.963
0.901
0.315
0.481
0.611
1.000
0.974
0.910
0.429
0.518
0.607
0.976
0.952
0.893
Dawson
5
4
3
0.752
0.782
0.809
0.877
0.815
0.765
0.759
0.796
0.815
0.810
0.744
0.718
0.745
0.768
0.804
0.940
0.881
0.810
Dawson-HHH
4
3
2
0.727
0.782
0.827
0.870
0.809
0.759
0.685
0.741
0.815
0.821
0.744
0.667
0.768
0.821
0.839
0.917
0.869
0.845
Hajizadeh
2
1.5
1
0.482
0.573
0.691
0.969
0.944
0.883
0.444
0.519
0.667
1.000
0.974
0.897
0.518
0.625
0.714
0.940
0.917
0.869
Diffey
3
2.5
2
0.845
0.880
0.880
0.790
0.784
0.741
0.833
0.852
0.852
0.744
0.744
0.718
0.857
0.907
0.907
0.833
0.821
0.762
Tronnier
4
3
2.5
0.791
0.818
0.827
0.815
0.790
0.722
0.778
0.796
0.815
0.769
0.718
0.667
0.804
0.839
0.839
0.857
0.857
0.774
Criteria
Algorithm
HHH
IHB
Dawson
Dawson-HHB
Hajizadeh
Diffey
Tronnier
Results
Percent
Difference
Sens
Overall
Spec
Sens
Dark
Spec
Sens
Light
Spec
40
30
20
30
20
15
20
10
5
15
10
5
30
20
10
20
15
10
30
20
15
0.873
0.901
0.945
0.809
0.840
0.830
0.591
0.818
0.882
0.664
0.755
0.836
0.784
0.832
0.847
0.845
0.873
0.891
0.736
0.845
0.861
0.887
0.813
0.722
0.732
0.703
0.638
0.932
0.796
0.660
0.932
0.840
0.747
0.773
0.680
0.600
0.741
0.679
0.636
0.759
0.698
0.648
0.907
0.926
1.000
0.789
0.842
0.842
0.481
0.741
0.833
0.574
0.685
0.815
0.826
0.870
0.783
0.796
0.833
0.870
0.611
0.796
0.827
0.910
0.833
0.744
0.611
0.704
0.593
0.936
0.782
0.654
0.923
0.821
0.744
0.712
0.636
0.591
0.769
0.705
0.641
0.782
0.705
0.679
0.833
0.872
0.893
0.821
0.839
0.821
0.696
0.893
0.929
0.750
0.821
0.857
0.750
0.800
0.904
0.893
0.911
0.911
0.857
0.893
0.893
0.861
0.792
0.702
0.810
0.702
0.667
0.929
0.810
0.667
0.940
0.857
0.750
0.821
0.714
0.607
0.714
0.655
0.631
0.738
0.690
0.619
22
Journal of Rehabilitation Research and Development Vol. 38 No. 1 2001
REFERENCES
1. Feather JW, Ryatt KS, Dawson JB. Reflectance spectrophotometric quantification of skin colour changes induced by topical corticosteroid preparations. Br J Dermatol 1982;106:436–40.
2. Wan S, Parrish JA, Jaenicke KF. Quantitative evaluation of ultraviolet induced erythema. Photochem Photobio 1983;37:643–8.
3. Hagisawa S, Ferguson-Pell M, Cardi M, Miller SD. Assessment of
skin blood content and oxygenation in spinal cord injured subjects
during reactive hyperemia. J Rehabil Res Dev 1994;31(1):1–14.
4. Takiwaki H, Serup, J. Measurement of erythema and melanin
indices. In: Serup J, Jemee GBE (editors). Handbook of non-invasive methods and the skin. Boca Raton, FL: CRC Press; 1995. p
377–97.
5. Dawson JB, Barker DJ, Ellis DJ, Grassam E, Cotterill JA, Fisher
GW, Feather JW. A theoretical and experimental study of light
absorption and scattering by in vivo skin. Phys Med Biol
1980;25(4)695–709.
6. Ferguson-Pell M, Hagisawa S. An empirical technique to compensate
for melanin when monitoring skin microcirculation using reflectance
spectrophotometry. Med Eng Physics 1995;17(2):104–10.
7. Feather JW, Hajizadeh-Saffar M, Leslie G, Dawson JB. A portable
scanning reflectance spectrophotometer using visible wavelengths
for the rapid measurement of skin pigments. Phys Med Biol
1989;34(7):807–20.
8. Hajizadeh-Saffar M, Feather JW, Dawson JB. An investigation of
factors affecting the accuracy of in vivo measurement of skin pigments by reflectance spectrophotometry. Phys Med Biol
1990;35(9):1301–15.
9. Diffey BL, Farr PM. Quantitative aspects of ultraviolet erythema.
Clin Phys Physiol Meas 1991;12(4):311–25.
10. Tronnier H. Evaluation and measurement of ultraviolet erythema.
In: F Urbach (editor). Biologic effects of ultraviolet radiation.
Oxford: Pergaman Press; 1969. p. 255–66.
11. Ferguson-Pell MW, Hagisawa S, Masiello RD. A skin indentation
system using a pneumatic bellows. J Rehabil Res Dev
1994;31(1):15–9.
12. Young KA. A study of the accuracy of pulse oximetry among varying skin pigment populations [master’s thesis]. Long Beach (CA):
California State Univ.; May 1994.
13. Landis EM. Micro-injection studies of capillary blood pressure in
human skin. Heart 1930;15:209–28.
Submitted for publication March 3, 2000. Accepted in
revised form June 22, 2000.