- Email: [email protected]
Muscle oxygenation measurement in humans by noninvasive optical spectroscopy and Locally Weighted Regression
Accepted Manuscript Title: Muscle Oxygenation Measurement in Humans by Noninvasive Optical Spectroscopy and Locally Weighted Regression Author: Lorilee S.L. Arakaki Kenneth A. Schenkman Wayne A. Ciesielski Jeremy M. Shaver PII: DOI: Reference:
S0003-2670(13)00652-1 http://dx.doi.org/doi:10.1016/j.aca.2013.05.003 ACA 232574
To appear in:
Analytica Chimica Acta
Received date: Revised date: Accepted date:
7-12-2012 29-4-2013 3-5-2013
Please cite this article as: L.S.L. Arakaki, K.A. Schenkman, W.A. Ciesielski, J.M. Shaver, Muscle Oxygenation Measurement in Humans by Noninvasive Optical Spectroscopy and Locally Weighted Regression, Analytica Chimica Acta (2013), http://dx.doi.org/10.1016/j.aca.2013.05.003 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
1 2
THIS IS THE HIGHLIGHTED COPY. ALL REVISIONS TO THE ORIGINAL MANUSCRIPT ARE UNDERLINED AND ARE IN RED.
3 Muscle Oxygenation Measurement in Humans by Noninvasive Optical Spectroscopy and
5
Locally Weighted Regression
6
Lorilee S. L. Arakaki,a,* Kenneth A. Schenkman,a,b Wayne A. Ciesielski,a and Jeremy M. Shaverc
cr
ip t
4
7
us
8 9
11
a
12
University of Washington
13
Box 356320
14
Seattle, WA 98195
15
USA
an
10
te
d
M
Department of Pediatrics
Ac ce p
16 17
b
18
University of Washington
19
Box 356540
20
Seattle, WA 98195
21
USA
Department of Anesthesiology
22 23
c
24
3905 West Eaglerock Dr.
25
Wenatchee, WA 98801
26
USA
Eigenvector Research, Inc.
Page 1 of 25
*Corresponding Author
28
Department of Pediatrics
29
University of Washington
30
Box 356320
31
Seattle, WA 98195
32
USA
33
Tel: 1-206-543-3373
34
Email: [email protected]
us
cr
ip t
27
an
35 36
M
37 Email addresses for co-authors:
39
Kenneth A. Schenkman: [email protected]
40
Wayne A. Ciesielski: [email protected]
41
Jeremy M. Shaver: [email protected]
te
Ac ce p
42
d
38
Page 2 of 25
42 Highlights
44
We propose a new method for measuring muscle oxygenation (Mox) from optical spectra.
45
Optical reflectance spectra were measured from the hand in the visible-NIR region (540-800 nm).
46
Locally Weighted Regression (LWR) with Partial Least Squares (PLS) was implemented.
47
Muscle oxygenation using LWR-PLS was more accurate than standard PLS.
48
Multivariate Curve Resolution can estimate Mox under controlled experimental conditions.
cr
us
49
Abstract
We have developed a method to make real-time, continuous, noninvasive measurements of
M
52
an
50 51
ip t
43
muscle oxygenation (Mox) from the surface of the skin. A key development was measurement in both
54
the visible and near infrared (NIR) regions. Measurement of both oxygenated and deoxygenated
55
myoglobin and hemoglobin resulted in a more accurate measurement of Mox than could be achieved
56
with measurement of only the deoxygenated components, as in traditional near-infrared spectroscopy
57
(NIRS). Using the second derivative with respect to wavelength reduced the effects of scattering on the
58
spectra and also made oxygenated and deoxygenated forms more distinguishable from each other.
59
Selecting spectral bands where oxygenated and deoxygenated forms absorb filtered out noise and
60
spectral features unrelated to Mox. NIR and visible bands were scaled relative to each other in order to
61
correct for errors introduced by normalization. Multivariate Curve Resolution (MCR) was used to
62
estimate Mox from spectra within each data set collected from healthy subjects. A Locally Weighted
63
Regression (LWR) model was built from calibration set spectra and associated Mox values from 20
64
subjects using 2562 spectra. LWR and Partial Least Squares (PLS) allow accurate measurement of Mox
65
despite variations in skin pigment or fat layer thickness in different subjects. The method estimated Mox
66
in five healthy subjects with an RMSE of 5.4%.
Ac ce p
te
d
53
Page 3 of 25
67 KEYWORDS
69
Near infrared spectroscopy, Visible spectroscopy, Locally weighted regression, Muscle oxygenation,
70
Tissue oxygenation, Hemoglobin
ip t
68
71 ABBREVIATIONS
73
Muscle oxygenation (Mox), Near infrared (NIR), Near-infrared spectroscopy (NIRS), Locally Weighted
74
Regression (LWR), Multivariate Curve Resolution (MCR), Partial Least Squares (PLS), total absorbance
75
summed across all wavelengths (Aobs), effective path length (B), total absorptivities summed across all
76
wavelengths of the oxygenated and deoxygenated species, respectively (Eo and Ed), concentrations of
77
oxygenated and deoxygenated species, respectively (Co and Cd), root mean squared error of cross-
78
validation (RMSECV), principal components (PCs), root mean square error (RMSE)
us
an
M
d te
80
Ac ce p
79
cr
72
Page 4 of 25
80 81 82
1. INTRODUCTION Optical spectroscopy is increasingly being used in the clinical setting to noninvasively monitor patients. The nonionizing radiation used and the potential for inexpensive, portable devices make
84
optical spectroscopy attractive for continuous monitoring in challenging clinical settings like
85
ambulances, emergency departments, intensive care units, and operating rooms.
cr
Hemoglobin is one of the most commonly measured biological chromophores. It is a protein in
us
86
ip t
83
blood that has four oxygen binding sites. Pulse oximetry is the standard of care in hospitals and uses two
88
wavelengths, one in the visible region and one in the infrared, to measure the oxygen saturation of
89
hemoglobin in arterial blood. While pulse oximetry is valuable in the assessment of oxygen availability in
90
the arterial blood, it does not yield information about whether enough oxygen is actually getting into
91
tissues to meet metabolic demands.
M
an
87
Because oxygen is essential for life, an enormous amount of medical treatment centers around
93
normalizing or stabilizing oxygen levels in the body. Despite significant advances in medical monitoring,
94
there currently is no device that continuously and quantitatively measures oxygen levels at the cellular
95
level, where oxygen is metabolized. Recently, there has been interest in optical monitoring of tissue
96
oxygenation in brain [1, 2] and muscle [2, 3]. Near-infrared spectroscopy (NIRS) devices typically
97
measure 2-6 wavelengths in the NIR region. Measurement of only a few wavelengths limits the ability to
98
sensitively measure tissue hemoglobin saturation. In a study of trauma patients, a commercially
99
available NIRS device that measures six wavelengths could identify severe hemorrhagic shock, but could
100
not distinguish mild and moderate shock from healthy controls [4]. Another NIRS device that measures
101
two NIR wavelengths has found increasing popularity as a trend monitor in operating rooms and
102
intensive care units, but its clinical value has been questioned [5, 6].
Ac ce p
te
d
92
Page 5 of 25
2 103
Our laboratory has measured myoglobin saturation [7-10] and hemoglobin saturation [11, 12] in vivo using Partial Least Squares (PLS) calibrated with spectra of solutions acquired in vitro. Myoglobin is
105
a protein found within muscle cells, and has one oxygen binding site. It functions as an intracellular store
106
and carrier for oxygen [13]. Our approach using PLS with full spectra in the visible and NIR regions
107
allowed accurate and separate measurements of myoglobin and hemoglobin saturation despite the fact
108
that the two proteins have very similar and highly overlapping absorbance spectra. PLS calibration on in
109
vitro spectra is not practical for clinical monitoring because PLS estimates are typically only valid for
110
subject spectra containing physiological interferences very similar to those in the calibration set [14]. It
111
is impractical and perhaps impossible to make appropriate in vitro calibration sets when attempting to
112
make single, instantaneous muscle oxygenation (Mox) estimates on subjects randomly selected from a
113
heterogeneous population.
M
an
us
cr
ip t
104
The goal of the present study was to develop a spectral analysis method to yield accurate, real-
115
time measurements of Mox that could potentially be used in a clinical monitor. Mox is a measure of the
116
percentage of oxygenated myoglobin and hemoglobin out of the total myoglobin and hemoglobin in the
117
optical path. Individual saturations of myoglobin and hemoglobin were not determined. Mox has
118
contributions from venous, capillary, and arterial blood and myoglobin.
Ac ce p
te
d
114
119
Full spectra were acquired between 540 and 800 nm, allowing measurement of both
120
oxygenated and deoxygenated forms of myoglobin and hemoglobin. This is in contrast to NIRS devices
121
that measure only deoxygenated hemoglobin in the NIR. Preprocessing included second derivatives,
122
variable selection, scaling between visible and NIR regions to balance information content between the
123
two regions, and total signal normalization. Multivariate Curve Resolution (MCR) was used to estimate
124
Mox in twenty healthy subjects. Those Mox estimates were then used in the Locally Weighted
125
Regression (LWR) training set. LWR provides a means to utilize the stability and simplicity of a linear
126
regression model in the presence of non-linear, difficult-to-model interferences [15]. In the case of
Page 6 of 25
3 transcutaneous Mox estimation, the various physiological variations from subject to subject (e.g.,
128
variations in skin tone and intervening fat layer thickness) are the nonlinear effects to be minimized. For
129
each measurement of a subject for which Mox is to be estimated, LWR automatically selects the most
130
similar subset of the training set and builds a "local" PLS model to estimate Mox. This provides a PLS
131
model that is more valid than one based on in vitro spectra because it is built from human subjects with
132
similar physical characteristics, notably skin tone and body mass index. The method was validated on
133
five additional healthy subjects.
us
cr
ip t
127
135
2. MATERIALS AND METHODS
136
2.1 EXPERIMENTAL
2.1.1. Optical spectroscopy. The optical system is composed of an imaging spectrometer (JY
M
137
an
134
Horiba iHR320), a thermoelectrically-cooled CCD camera (JY Horiba Synapse, 256x1024 format), a
139
custom-designed LED light source (Innovations in Optics, Inc.), and a custom-designed fiber optic probe
140
(Fiberoptic Systems, Inc.). The 300-groove grating in the spectrometer was centered at 670 nm and the
141
wavelength resolution was approximately 0.25 nm/pixel.
te
Ac ce p
142
d
138
The fiber optic probe is bifurcated, with one branch connected to the light source and the other
143
branch coupled to the spectrometer entrance slit. The common end of the probe was placed on the
144
back of the subject’s hand over the thenar muscle group and was held in place with an elastic bandage.
145
Three probes were used, each with a constant separation between source and detection fibers of 13, 14,
146
or 15 mm. The largest probe that yielded spectra with adequate signal-to-noise ratios was used in each
147
individual subject. Hand spectra were referenced to a spectrum collected from 1% Intralipid to calculate
148
an approximate absorbance response.
149 150
2.1.2. Human subject protocol. With approval from the University of Washington Institutional Review Board, healthy controls between the ages of 18 and 80 were recruited. After consent was
Page 7 of 25
4 obtained, subjects were interviewed to determine whether they had cardiovascular or pulmonary
152
disease, diabetes, or a history of deep vein thrombosis. Subjects were excluded from the study if they
153
had one or more of these conditions. Baseline blood pressures and pulse rates were recorded, in
154
addition to height and weight. Twenty-five subjects were studied, with body mass indices of 24.2 ± 3.0.
155
ip t
151
Spectra (540-800 nm) were acquired at a rate of either 0.33 or 0.5 Hz during a 3-min period of rest while each subject breathed room air. An oxygen mask was then placed over the face for 3 min, and
157
the subject breathed 100% O2. A blood pressure cuff was inflated on the upper arm to a pressure that
158
was 60 mmHg above the subject’s systolic blood pressure or to 200 mmHg, whichever was less. The cuff
159
was inflated for 15 min, and the oxygen mask was removed approximately two minutes into the period
160
of arm ischemia. A recovery period of 5 min followed to ensure normalization of blood perfusion to the
161
arm. Spectral acquisition was continuous throughout the entire protocol.
us
an
M
162
cr
156
Consecutive spectra were averaged to improve signal-to-noise ratios. When the acquisition rate was 0.33 Hz, three spectra were averaged together, producing a new rate of 0.11 Hz. When the
164
acquisition rate was 0.5 Hz, five spectra were averaged together for a new rate of 0.1 Hz.
te
d
163
Spectra collected from the hand at rest while breathing 100% O2 represented the most
166
oxygenated condition that we could practically produce in the subjects (Fig.1). We assumed that these
167
spectra represented Mox values of 100% but we have not verified with an independent measure that
168
myoglobin and hemoglobin in the hand were fully saturated with oxygen. Spectra collected at the end of
169
the 15-min period of arm ischemia were assumed to represent Mox values of 0% (Fig. 1). In healthy
170
subjects, it has been observed that 15 min of arm ischemia is sufficient to achieve complete
171
deoxygenation of skeletal muscle and blood in the hand because myoglobin and hemoglobin saturations
172
plateau to a minimum value [12]. Spectra acquired during the period of arm ischemia represent
173
intermediate Mox values between 0 and 100%.
174
2.2. ANALYSIS
Ac ce p
165
Page 8 of 25
5 175 176
All analyses were performed using PLS_Toolbox Version 6.2 (Eigenvector Research, Inc.) running in Matlab R2009b (Mathworks, Inc.). To help remove baseline and background signals due to scattering phenomena, second
178
derivatives with respect to wavelength were taken using the Savitzky-Golay algorithm with a 65-point
179
window and a second order polynomial [16]. The visible wavelength region was restricted to 562.2 –
180
601.7 nm and the NIR region was restricted to 744.7 – 802.2 nm. These wavelength regions were
181
selected because they encompass the absorbances of the oxygenated and deoxygenated species,
182
respectively. The two regions were scaled relative to each other to correct for differences in absorptivity
183
of the species in the visible and infrared regions of the spectra. The magnitude of this adjustment was
184
determined using the method described in section 2.2.2. The two discontinuous wavelength regions
185
were then concatenated. Finally, to correct for changes in illumination and path length, as well as other
186
total-signal influences, each spectrum was normalized to its total summed absorbance.
cr
us
an
M
2.2.1. Estimation of Mox Reference Values by MCR. As no accurate primary method exists to
d
187
ip t
177
determine the actual Mox in human subjects at any given point in time, MCR [17, 18] was used to
189
estimate Mox during ischemia. Separate two-factor MCR models were calculated for each subject to
190
recover estimates of the oxygenated and deoxygenated component spectra and corresponding
191
contributions of each species at each point in time.
Ac ce p
te
188
192
Each model used a closure constraint to assure that the summation of the two recovered
193
contribution profiles would always be 100%. Assuming that the spectra collected just prior to cuff
194
inflation during breathing of pure oxygen are close to 100% Mox, the five spectra at this point in every
195
run were constrained to give a 100%:0% ratio of oxygenated to deoxygenated species. Likewise, the last
196
five spectra in each run prior to cuff release were constrained to give a 0%:100% ratio. The average
197
spectra from each of these time periods were used as the initial guesses for the pure-component
198
spectra in the MCR. Because the closure constraint requires that the two contribution profiles sum to
Page 9 of 25
6 199
100%, the recovered oxygenated species contribution profile is equivalent to the estimated Mox for the
200
subject. An example Mox profile is shown in Fig. 2.
201
It is important to note that this procedure to recover the Mox profiles is only an approximation of the actual Mox, which may be influenced by optical variations (such as skin-probe boundary changes).
203
However, it is also assumed that when all the subjects' results are combined during the LWR modeling
204
later, these effects will be diminished by an averaging of the pool of all subjects.
cr
ip t
202
2.2.2. Correction of Absorptivity Effects. Because there is always an ambiguity in the total
206
volume of muscle interrogated, the total observed absorbance is also ambiguous. The absorbance
207
observed for the deoxygenated species, for example, may increase due to decreased Mox, or due to an
208
increase in total interrogated volume. To correct for these effects, each spectrum is normalized to the
209
total absolute value of the observed signal (the sum of the absolute value) after taking the second
210
derivative and before MCR. This has two side effects: First, it means relative absorbances of the
211
oxygenated and deoxygenated species can be used to calculate Mox. Second, it introduces an
212
unintended nonlinearity between calculated Mox and true Mox due to the difference in absorptivity
213
between the oxygenated and deoxygenated species, as described below. Note that this effect is not a
214
nonlinearity between observed absorbances at different wavelengths (which would lead to a higher
215
multivariate rank of the set of spectra), but one simply between the calculated and true Mox.
217 218
an
M
d
te
Ac ce p
216
us
205
The total absorbance summed across all wavelengths, Aobs, can be written as a summation of the absorbances of the oxygenated and deoxygenated species: Aobs = EoBCo + EdBCd
(1)
219
where B is the effective path length, Eo and Ed are the total absorptivities summed across all
220
wavelengths of the oxygenated and deoxygenated species, respectively, and Co and Cd are the
221
concentrations of oxygenated and deoxygenated species, respectively. If the two absorptivity profiles
222
had the same integrated areas, Aobs would only be influenced by changes in path length or total
Page 10 of 25
7 223
myoglobin + hemoglobin concentration (Co + Cd) and dividing each spectrum by Aobs would correct for
224
changes in those properties. However, because Eo and Ed are not equal, the total summed absorbance,
225
Aobs, will vary with Mox even in the absence of changes in path length or Co + Cd. Fig. 3 illustrates the problem that would occur with normalization if there were a 4:1 ratio
ip t
226
between Eo and Ed. The dotted lines represent "true" oxygenated and deoxygenated species
228
concentrations (in relative %) as Mox decreases linearly from left to right. The solid lines show the
229
concentrations that would be calculated from normalized spectra. At 50% true Mox, the observed Mox
230
would be 80%, reflecting the 4:1 contributions to Aobs from Eo and Ed when Co = Cd. At 100% Mox, the
231
summed absorbance is dominated by the oxygenated species absorbance (Aobs = EoBC), while at 0% Mox,
232
the summed absorbance is dominated by the deoxygenated species (Aobs = EdBC). Aobs at 100% Mox
233
would be four times greater than Aobs at 0% Mox. Thus, as Mox varies, Aobs is not a constant but is
234
changing as a proportion of the absorptivity of the oxygenated or deoxygenated species. The effect of
235
the non-constant Aobs is that observed Mox would be consistently overestimated in this example.
us
an
M
d
One approach to correct for this effect is to scale the visible region of the spectrum (dominated
te
236
cr
227
by the oxygenated species) relative to the NIR region (dominated by the deoxygenated species) by a
238
factor corresponding to the ratio between Eo and Ed. Estimating an appropriate scaling factor was
239
achieved by taking advantage of the fact that when there are no major path length effects (such as in
240
the highly-controlled ischemia trials), normalization has little effect except to introduce the nonlinearity.
241
Once the scaling factor is correct, normalization should have almost no effect on Mox values recovered
242
using MCR. By adjusting the NIR:Visible scaling factor and then examining Mox profiles recovered with
243
and without normalization, we identified a factor at which the profiles were not significantly changed by
244
normalization. This was considered the "optimal scaling factor" for the given subject. The 50% Mox point
245
was used as a reference to indicate how similar the profiles were with and without normalization.
Ac ce p
237
Page 11 of 25
8 246
Figs. 4A and 4B show an example of the change in Mox profile resulting from normalization
247
before a scaling factor was applied. In Figs. 4C and 4D, the 50% Mox points of unnormalized and
248
normalized data sets were identical when a scaling factor of 5.3 was used. This process was repeated for all 20 training subjects and the median optimal scaling factor of
ip t
249
4.7 was selected for all subsequent spectral adjustments. Each training subject's spectra were corrected
251
by the median optimal scaling factor and the Mox profiles were estimated using the MCR procedure
252
described earlier. Finding a single factor to use for all subjects provides for the most general correction
253
of absorptivity and possible optical effects without over-correcting for a single subject. It also provides a
254
means to analyze subjects for which the normalized and non-normalized results were too severely
255
different (probably due to actual path length or optical effects) and enables use of the models for
256
application to future clinical subjects for which we will have only a single point, and not an entire
257
ischemic profile.
us
an
M
2.2.3. Locally Weighted Regression Modeling. Given the estimated Mox values and the
d
258
cr
250
individual spectra for all subjects as a whole, an LWR model was created that used a two principal
260
component PCA model to select training spectra. For any given Mox estimate to be made, 200 training
261
spectra were selected and a one latent variable PLS model was calculated and used for predictions. An
262
attractive feature of such an empirical, inverse least squares PLS model is that having relatively few
263
constraints allows flexibility in correcting for interferences. In contrast, a classical least squares method
264
like MCR will have a greater number of constraints that will likely degrade prediction accuracy.
Ac ce p
265
te
259
The training spectra provided to the LWR model were pre-processed similarly to that described
266
for the MCR estimation procedure (2nd derivative, 4.7 NIR:Vis scaling, area normalization) but the
267
spectra and estimated Mox values were also mean-centered to allow for an intercept in the linear
268
regression. The number of principal components, number of training spectra selected, and number of
269
PLS latent variables to be used were all selected using cross-validation in which the entire set of spectra
Page 12 of 25
9 from a given subject were left out of the training set, and the remaining 19 subjects were used to
271
estimate the left-out subject's Mox. The error in these Mox estimations was summarized in the root
272
mean squared error of cross-validation (RMSECV) which is the square root of the average error of
273
prediction for all samples when they were left out of the training set [14]. This was repeated for various
274
settings of the LWR algorithm with general agreement that the selected LWR parameters gave stable
275
and reliable results.
cr
2.2.4. Partial Least Squares Modeling. For comparison to the LWR models, an identical
us
276
ip t
270
modeling effort was made using standard (non-locally-weighted) PLS modeling. For these models, the
278
identical calibration data were analyzed with a standard PLS algorithm using cross-validation to
279
determine the number of PLS latent variables. The best smallest cross-validation error was observed at 2
280
latent variables.
M
281
an
277
2.2.5. Quality of Spectral Fit Parameters. In both the PLS and LWR models, it is important to assess how well the model fits a given spectrum when making a Mox prediction for that spectrum. For
283
factor-based models such as PLS and LWR, this can be done using the Hotelling’s T2 and Q residuals [14].
284
Generically, these models decompose a data matrix, X (in which rows are spectra), into scores, T, and
285
loadings, P, and spectral residuals, R, according to the equation:
Ac ce p
te
d
282
286
X = TPT + R
287
Hotelling’s T2 is calculated as a weighted sum of the scores and describes whether the features
(2)
288
described by the loadings of the model are present in similar scale to what was observed in the
289
calibration data. The Q residuals are calculated as the sum of the squared residuals in each row of the R
290
matrix [14]. When making a prediction for a new sample, a new Hotelling’s T2 value and Q residual value
291
are calculated and compared to the values obtained for the calibration data. If a new value is
292
substantially different from the calibration data, it is inferred that the new spectrum contains some
Page 13 of 25
10 293
feature(s) which were not observed in the calibration data and, thus, the predictions for Mox should not
294
be used.
295
297
3. RESULTS AND DISCUSSION
ip t
296
3.1. Cross-Validation. While building the models, cross-validation was used as described earlier. For the LWR model, cross-validation indicated very little change as a function of either the number of
299
principal components (PCs) in the model or the number of samples selected for each local model.
300
Models were examined with 1 to 6 PCs, 100, 150, or 200 selected samples, and 1 to 3 latent variables in
301
the local PLS model. The RMSECV ranged from 6.5% at 2 PCs and 200 samples selected for each local
302
model up to 8.1% at 6 PCs and 100 samples selected. In comparison, the best standard PLS model (not
303
LWR-based model) provided an RMSECV of 9.3% obtained with 1 latent variable. All LWR and PLS
304
models with more than 1 latent variable gave RMSECVs above 9.3%. In all cases, the subjects in the
305
calibration set displayed Hotelling’s T2 and Q residuals that indicated there were no unusual subjects or
306
samples in the calibration set.
us
an
M
d
te
3.2. Validation Subject Results. In order to validate the model, Mox measurements were made
Ac ce p
307
cr
298
308
from five healthy subjects that were not included in the model. These subjects underwent the same
309
protocol as the 20 subjects included in the model. Fig. 5 shows that the LWR model produced accurate
310
Mox measurements compared to MCR estimates of Mox obtained from each subject’s data set. The
311
best-fit line had a slope of 0.99 and a y-intercept of -2.2, very similar to the line of unity. The root mean
312
square error (RMSE) for all 654 spectra was 5.4%.
313
For comparison, the results for the standard PLS model applied to the same five test subjects are
314
shown in Fig. 6. Because the PLS model is required to take into account all subjects equally, the model
315
must be a “consensus,” resulting in a per-subject bias. Three of the five subjects show notable
Page 14 of 25
11 316
systematic offset from the MCR estimates. The RMSE for the test subjects using the PLS model was
317
9.6%.
318
Finally, an additional LWR model was built with the wavelength range limited to only the NIR portion of the spectrum (744.7 – 802.2 nm). In this model, the predictions showed little to no trend with
320
Mox and produced an RMSE of 29.3% (Fig. 7). Such a poor result is expected considering that removing
321
the visible portion of the spectrum leaves no ability to correct for path length and absorbance effects. It
322
demonstrates the need for obtaining both the visible and NIR portions together for the Mox evaluation.
323
Interestingly, these poor results are probably overly optimistic considering that reference Mox values
324
were obtained from MCR using the full wavelength range. If the MCR step was similarly limited in
325
wavelength window, no sensible reference values could even be obtained for the calibration samples.
cr
us
an
3.3. Application to Trauma Patients. Mox was measured in a preliminary study of trauma
M
326
ip t
319
patients admitted to Harborview Medical Center in Seattle, WA. Spectra were acquired for 15 min as
328
soon as possible upon admission to the emergency department. The LWR model was applied to the
329
spectra from each patient. Fig. 8 shows the Mox time course in one patient. Mox has a good correlation
330
with blood pressure; it provides a continuous and real-time indication of tissue perfusion. The sharp
331
increases in Mox at approximately t = 2 and 8 min correlate with intravenous saline infusions given for
332
low blood pressure that temporarily improved tissue perfusion. In contrast, arterial blood saturation
333
measured by pulse oximetry was not sensitive to changes in tissue perfusion.
335 336
te
Ac ce p
334
d
327
4. CONCLUSIONS
The results for both the validation subjects and a trauma patient indicate that the use of joined
337
visible and NIR spectra can provide Mox estimates better than what has been previously reported for
338
single- or double-wavelength NIR methods. In addition, these results show that the LWR models are
339
capable of making reasonable estimates of Mox, even when presented with the limited range of
Page 15 of 25
12 340
physiological variability provided in this small calibration set. The accuracy of these predictions should
341
only improve when additional calibration subjects, representing a wider variety of skin tones and body
342
mass indices, are added to the calibration set. Further improvements in model performance would also be expected by using a modified
ip t
343
method for training sample selection. In this work, the sample-to-sample distance in the PCA score
345
space was used. This PCA model captured the variance associated with both the subject-to-subject
346
variability (which is most critical to correct) and the varying Mox caused by ischemia in individual
347
subjects. Thus, samples were selected both on the similarity of the subject as well as the similarity of the
348
extent of ischemia to test spectra. The predictions made from the ordinary PLS model show that the
349
calibration and test subjects were generally in error by a constant slope and/or offset across all Mox
350
values. From this, we would anticipate that a modified sample selection method which focuses more on
351
subject-to-subject variability (for example, discarding all spectra for individuals which are dissimilar and
352
keeping all spectra for individuals which are similar) would tend to create models that capture the
353
specific slope and offset scaling relevant for the tested individual. Such models may provide better
354
predictions than our current model.
us
an
M
d
te
Ac ce p
355
cr
344
This work also demonstrates that the use of MCR with simple equality and closure constraints
356
provides a reasonable inference of Mox in the absence of any true primary reference. This use of a
357
carefully controlled experimental protocol and inferential estimates of Mox can be used to easily
358
produce additional calibration data for the models.
359 360 361
ACKNOWLEDGMENTS This work was funded in part by grants from the Coulter Foundation, the Life Sciences Discovery
362
Fund and NIH 1R21EB009099. None of these granting agencies played any role in the study design or in
363
the collection, analysis or interpretation of the data.
Page 16 of 25
13 364 REFERENCES
366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391
[1] M. Smith, Philos Transact A Math Phys Eng Sci 369 (2011) 4452-4469. [2] J.M. Murkin, M. Arango, Br J Anaesth 103 Suppl 1 (2009) i3-13. [3] T.W. Scheeren, P. Schober, L.A. Schwarte, J Clin Monit Comput 26 (2012) 279-287. [4] B.A. Crookes, S.M. Cohn, S. Bloch, J. Amortegui, R. Manning, P. Li, M.S. Proctor, A. Hallal, L.H. Blackbourne, R. Benjamin, D. Soffer, F. Habib, C.I. Schulman, R. Duncan, K.G. Proctor, J Trauma 58 (2005) 806-813; discussion 813-806. [5] J.C. Hirsch, J.R. Charpie, R.G. Ohye, J.G. Gurney, Semin Thorac Cardiovasc Surg Pediatr Card Surg Annu 13 (2010) 51-54. [6] D. Highton, C. Elwell, M. Smith, Curr Opin Anaesthesiol 23 (2010) 576-581. [7] G.A. Cohen, L.C. Permut, L.S. Arakaki, W.A. Ciesielski, D.M. McMullan, A.R. Parrish, K.A. Schenkman, ASAIO J 57 (2011) 314-317. [8] K.A. Schenkman, D.R. Marble, D.H. Burns, E.O. Feigl, Appl Spectrosc 53 (1999) 332-338. [9] L.S.L. Arakaki, W.A. Ciesielski, B.D. Thackray, E.O. Feigl, K.A. Schenkman, Appl Spectrosc 64 (2010) 973-979. [10] L.S.L. Arakaki, D.H. Burns, M.J. Kushmerick, Appl Spectrosc 61 (2007) 978-985. [11] D.J. Marcinek, W.A. Ciesielski, K.E. Conley, K.A. Schenkman, Am J Physiol Heart Circ Physiol 285 (2003) H1900-H1908. [12] C.E. Amara, D.J. Marcinek, E.G. Shankland, K.A. Schenkman, L.S. Arakaki, K.E. Conley, Methods 46 (2008) 312-318. [13] J.B. Wittenberg, B.A. Wittenberg, J Exp Biol 206 (2003) 2011-2020. [14] H. Martens, T. Naes, Multivariate Calibration, John Wiley & Sons, Chichester, 1989. [15] T. Naes, T. Isaksson, B. Kowalski, Anal Chem 62 (1990) 664-673. [16] A. Savitzky, M.J.E. Golay, Anal Chem 36 (1964) 1627. [17] M.H. Van Benthem, M.R. Keenan, D.M. Haaland, J Chemom 16 (2002) 613-622. [18] P.J. Gemperline, E. Cash, Anal Chem 75 (2003) 4236-4243.
cr
us
an
M
d
te
Ac ce p
392
ip t
365
Page 17 of 25
14
us
cr
ip t
392
an
393
Figure 1. Spectra acquired from a healthy human subject at extremes of muscle oxygenation. The solid
395
spectrum was acquired after 3 min of 100% O2 breathing. The dashed spectrum was acquired after 15
396
min of arm ischemia. Oxymyoglobin and oxyhemoglobin absorb near 580 nm in the visible region;
397
deoxymyoglobin and deoxyhemoglobin absorb near 760 nm in the near infrared region. The visible and
398
NIR bands that were used to estimate Mox are indicated by the sets of vertical dashed lines. In the
399
wavelength region between 610 and 730 nm, the detector was saturated. These spectral values do not
400
represent measurements of tissue absorbance and are not used in the spectral analysis.
d
te
Ac ce p
401
M
394
Page 18 of 25
15 401
an
us
cr
ip t
402
403
Figure 2. Muscle oxygenation (Mox) in a healthy human subject estimated by MCR. The subject
405
breathed air for 3 min, then 100% O2 for 3 min, and then underwent a 15-min period of arm ischemia.
M
404
Ac ce p
te
d
406
Page 19 of 25
16 406
an
us
cr
ip t
407
408
Figure 3. Normalization bias effect. Since the integrated areas of absorptivity profiles of oxygenated and
410
deoxygenated species are not equal, normalization introduces an unintended nonlinearity to recovered
411
relative oxygenated and deoxygenated concentrations. Dotted lines represent true relative
412
concentrations and by definition, true Mox is equal to “True Oxygenated.” True Mox decreases from left
413
to right, as during ischemia. If the ratio of the integrated areas of oxygenated to deoxygenated
414
absorptivity profiles were 4:1, the observed relative concentrations would be those shown by the solid
415
lines. The observed Mox, equal to “Observed Oxygenated,” would be consistently overestimated.
d
te
Ac ce p
416
M
409
Page 20 of 25
17 416 417 418
ip t
419 420
cr
421
Ac ce p
te
d
M
an
us
422 423 424 425 426
427 428 429 430 431 432 433 434 435 436 437 438 439 440
Page 21 of 25
18 Figure 4. Scaling correction for absorptivity effects. A) MCR results without (left) and with (right)
442
normalization show that the 50% Mox point is affected by normalization. B) Concatenated second-
443
derivative spectra. The visible region comprises 562-601 nm and the NIR region comprises 744-802 nm.
444
C) Up-weighting the NIR region where deoxygenated species have predominant intensity corrects for
445
the normalization effect. This example shows results when the NIR region was up-scaled by a factor of
446
5.3. The optimal scaling factor was determined by adjusting the scaling until the 50% Mox points for
447
unnormalized (left) and normalized (right) cases were equal. D) Concatenated second-derivative spectra
448
with the scaling, without and with normalization.
us
cr
ip t
441
Ac ce p
te
d
M
an
449
Page 22 of 25
19
us
cr
ip t
449
an
450
Figure 5. Accuracy of muscle oxygenation (Mox) measurements by LWR in five healthy human subjects.
452
Mox was measured accurately across its range from 0% to 100%.
M
451
Ac ce p
455
te
454
d
453
456 457 458 459 460 461 462 463
464
Figure 6. Accuracy of muscle oxygenation (Mox) measurements by PLS in five healthy human subjects.
465
Most subjects show a systematic bias at all Mox levels.
Page 23 of 25
us
cr
ip t
20
466
Figure 7. Muscle oxygenation (Mox) measurements by LWR in five healthy human subjects when only
468
the NIR portion of the spectra is used. Accuracy is very poor. The advantage of including the visible
469
region in the analysis is apparent when these results are compared to those in Figure 5.
M
an
467
470
d
471
Ac ce p
te
472 473 474 475 476 477 478 479
480
Figure 8. Muscle oxygenation (Mox) is indicative of tissue perfusion. Mox was measured in a trauma
481
patient who received saline infusions at t = 2 and 8 min (indicated by arrows), causing increases in blood
482
pressure (BP) and corresponding increases in Mox. Arterial saturation measured by pulse oximetry was
483
not sensitive to changes in tissue perfusion.
Page 24 of 25
ce Ac
Page 25 of 25
S0003-2670(13)00652-1 http://dx.doi.org/doi:10.1016/j.aca.2013.05.003 ACA 232574
To appear in:
Analytica Chimica Acta
Received date: Revised date: Accepted date:
7-12-2012 29-4-2013 3-5-2013
Please cite this article as: L.S.L. Arakaki, K.A. Schenkman, W.A. Ciesielski, J.M. Shaver, Muscle Oxygenation Measurement in Humans by Noninvasive Optical Spectroscopy and Locally Weighted Regression, Analytica Chimica Acta (2013), http://dx.doi.org/10.1016/j.aca.2013.05.003 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
1 2
THIS IS THE HIGHLIGHTED COPY. ALL REVISIONS TO THE ORIGINAL MANUSCRIPT ARE UNDERLINED AND ARE IN RED.
3 Muscle Oxygenation Measurement in Humans by Noninvasive Optical Spectroscopy and
5
Locally Weighted Regression
6
Lorilee S. L. Arakaki,a,* Kenneth A. Schenkman,a,b Wayne A. Ciesielski,a and Jeremy M. Shaverc
cr
ip t
4
7
us
8 9
11
a
12
University of Washington
13
Box 356320
14
Seattle, WA 98195
15
USA
an
10
te
d
M
Department of Pediatrics
Ac ce p
16 17
b
18
University of Washington
19
Box 356540
20
Seattle, WA 98195
21
USA
Department of Anesthesiology
22 23
c
24
3905 West Eaglerock Dr.
25
Wenatchee, WA 98801
26
USA
Eigenvector Research, Inc.
Page 1 of 25
*Corresponding Author
28
Department of Pediatrics
29
University of Washington
30
Box 356320
31
Seattle, WA 98195
32
USA
33
Tel: 1-206-543-3373
34
Email: [email protected]
us
cr
ip t
27
an
35 36
M
37 Email addresses for co-authors:
39
Kenneth A. Schenkman: [email protected]
40
Wayne A. Ciesielski: [email protected]
41
Jeremy M. Shaver: [email protected]
te
Ac ce p
42
d
38
Page 2 of 25
42 Highlights
44
We propose a new method for measuring muscle oxygenation (Mox) from optical spectra.
45
Optical reflectance spectra were measured from the hand in the visible-NIR region (540-800 nm).
46
Locally Weighted Regression (LWR) with Partial Least Squares (PLS) was implemented.
47
Muscle oxygenation using LWR-PLS was more accurate than standard PLS.
48
Multivariate Curve Resolution can estimate Mox under controlled experimental conditions.
cr
us
49
Abstract
We have developed a method to make real-time, continuous, noninvasive measurements of
M
52
an
50 51
ip t
43
muscle oxygenation (Mox) from the surface of the skin. A key development was measurement in both
54
the visible and near infrared (NIR) regions. Measurement of both oxygenated and deoxygenated
55
myoglobin and hemoglobin resulted in a more accurate measurement of Mox than could be achieved
56
with measurement of only the deoxygenated components, as in traditional near-infrared spectroscopy
57
(NIRS). Using the second derivative with respect to wavelength reduced the effects of scattering on the
58
spectra and also made oxygenated and deoxygenated forms more distinguishable from each other.
59
Selecting spectral bands where oxygenated and deoxygenated forms absorb filtered out noise and
60
spectral features unrelated to Mox. NIR and visible bands were scaled relative to each other in order to
61
correct for errors introduced by normalization. Multivariate Curve Resolution (MCR) was used to
62
estimate Mox from spectra within each data set collected from healthy subjects. A Locally Weighted
63
Regression (LWR) model was built from calibration set spectra and associated Mox values from 20
64
subjects using 2562 spectra. LWR and Partial Least Squares (PLS) allow accurate measurement of Mox
65
despite variations in skin pigment or fat layer thickness in different subjects. The method estimated Mox
66
in five healthy subjects with an RMSE of 5.4%.
Ac ce p
te
d
53
Page 3 of 25
67 KEYWORDS
69
Near infrared spectroscopy, Visible spectroscopy, Locally weighted regression, Muscle oxygenation,
70
Tissue oxygenation, Hemoglobin
ip t
68
71 ABBREVIATIONS
73
Muscle oxygenation (Mox), Near infrared (NIR), Near-infrared spectroscopy (NIRS), Locally Weighted
74
Regression (LWR), Multivariate Curve Resolution (MCR), Partial Least Squares (PLS), total absorbance
75
summed across all wavelengths (Aobs), effective path length (B), total absorptivities summed across all
76
wavelengths of the oxygenated and deoxygenated species, respectively (Eo and Ed), concentrations of
77
oxygenated and deoxygenated species, respectively (Co and Cd), root mean squared error of cross-
78
validation (RMSECV), principal components (PCs), root mean square error (RMSE)
us
an
M
d te
80
Ac ce p
79
cr
72
Page 4 of 25
80 81 82
1. INTRODUCTION Optical spectroscopy is increasingly being used in the clinical setting to noninvasively monitor patients. The nonionizing radiation used and the potential for inexpensive, portable devices make
84
optical spectroscopy attractive for continuous monitoring in challenging clinical settings like
85
ambulances, emergency departments, intensive care units, and operating rooms.
cr
Hemoglobin is one of the most commonly measured biological chromophores. It is a protein in
us
86
ip t
83
blood that has four oxygen binding sites. Pulse oximetry is the standard of care in hospitals and uses two
88
wavelengths, one in the visible region and one in the infrared, to measure the oxygen saturation of
89
hemoglobin in arterial blood. While pulse oximetry is valuable in the assessment of oxygen availability in
90
the arterial blood, it does not yield information about whether enough oxygen is actually getting into
91
tissues to meet metabolic demands.
M
an
87
Because oxygen is essential for life, an enormous amount of medical treatment centers around
93
normalizing or stabilizing oxygen levels in the body. Despite significant advances in medical monitoring,
94
there currently is no device that continuously and quantitatively measures oxygen levels at the cellular
95
level, where oxygen is metabolized. Recently, there has been interest in optical monitoring of tissue
96
oxygenation in brain [1, 2] and muscle [2, 3]. Near-infrared spectroscopy (NIRS) devices typically
97
measure 2-6 wavelengths in the NIR region. Measurement of only a few wavelengths limits the ability to
98
sensitively measure tissue hemoglobin saturation. In a study of trauma patients, a commercially
99
available NIRS device that measures six wavelengths could identify severe hemorrhagic shock, but could
100
not distinguish mild and moderate shock from healthy controls [4]. Another NIRS device that measures
101
two NIR wavelengths has found increasing popularity as a trend monitor in operating rooms and
102
intensive care units, but its clinical value has been questioned [5, 6].
Ac ce p
te
d
92
Page 5 of 25
2 103
Our laboratory has measured myoglobin saturation [7-10] and hemoglobin saturation [11, 12] in vivo using Partial Least Squares (PLS) calibrated with spectra of solutions acquired in vitro. Myoglobin is
105
a protein found within muscle cells, and has one oxygen binding site. It functions as an intracellular store
106
and carrier for oxygen [13]. Our approach using PLS with full spectra in the visible and NIR regions
107
allowed accurate and separate measurements of myoglobin and hemoglobin saturation despite the fact
108
that the two proteins have very similar and highly overlapping absorbance spectra. PLS calibration on in
109
vitro spectra is not practical for clinical monitoring because PLS estimates are typically only valid for
110
subject spectra containing physiological interferences very similar to those in the calibration set [14]. It
111
is impractical and perhaps impossible to make appropriate in vitro calibration sets when attempting to
112
make single, instantaneous muscle oxygenation (Mox) estimates on subjects randomly selected from a
113
heterogeneous population.
M
an
us
cr
ip t
104
The goal of the present study was to develop a spectral analysis method to yield accurate, real-
115
time measurements of Mox that could potentially be used in a clinical monitor. Mox is a measure of the
116
percentage of oxygenated myoglobin and hemoglobin out of the total myoglobin and hemoglobin in the
117
optical path. Individual saturations of myoglobin and hemoglobin were not determined. Mox has
118
contributions from venous, capillary, and arterial blood and myoglobin.
Ac ce p
te
d
114
119
Full spectra were acquired between 540 and 800 nm, allowing measurement of both
120
oxygenated and deoxygenated forms of myoglobin and hemoglobin. This is in contrast to NIRS devices
121
that measure only deoxygenated hemoglobin in the NIR. Preprocessing included second derivatives,
122
variable selection, scaling between visible and NIR regions to balance information content between the
123
two regions, and total signal normalization. Multivariate Curve Resolution (MCR) was used to estimate
124
Mox in twenty healthy subjects. Those Mox estimates were then used in the Locally Weighted
125
Regression (LWR) training set. LWR provides a means to utilize the stability and simplicity of a linear
126
regression model in the presence of non-linear, difficult-to-model interferences [15]. In the case of
Page 6 of 25
3 transcutaneous Mox estimation, the various physiological variations from subject to subject (e.g.,
128
variations in skin tone and intervening fat layer thickness) are the nonlinear effects to be minimized. For
129
each measurement of a subject for which Mox is to be estimated, LWR automatically selects the most
130
similar subset of the training set and builds a "local" PLS model to estimate Mox. This provides a PLS
131
model that is more valid than one based on in vitro spectra because it is built from human subjects with
132
similar physical characteristics, notably skin tone and body mass index. The method was validated on
133
five additional healthy subjects.
us
cr
ip t
127
135
2. MATERIALS AND METHODS
136
2.1 EXPERIMENTAL
2.1.1. Optical spectroscopy. The optical system is composed of an imaging spectrometer (JY
M
137
an
134
Horiba iHR320), a thermoelectrically-cooled CCD camera (JY Horiba Synapse, 256x1024 format), a
139
custom-designed LED light source (Innovations in Optics, Inc.), and a custom-designed fiber optic probe
140
(Fiberoptic Systems, Inc.). The 300-groove grating in the spectrometer was centered at 670 nm and the
141
wavelength resolution was approximately 0.25 nm/pixel.
te
Ac ce p
142
d
138
The fiber optic probe is bifurcated, with one branch connected to the light source and the other
143
branch coupled to the spectrometer entrance slit. The common end of the probe was placed on the
144
back of the subject’s hand over the thenar muscle group and was held in place with an elastic bandage.
145
Three probes were used, each with a constant separation between source and detection fibers of 13, 14,
146
or 15 mm. The largest probe that yielded spectra with adequate signal-to-noise ratios was used in each
147
individual subject. Hand spectra were referenced to a spectrum collected from 1% Intralipid to calculate
148
an approximate absorbance response.
149 150
2.1.2. Human subject protocol. With approval from the University of Washington Institutional Review Board, healthy controls between the ages of 18 and 80 were recruited. After consent was
Page 7 of 25
4 obtained, subjects were interviewed to determine whether they had cardiovascular or pulmonary
152
disease, diabetes, or a history of deep vein thrombosis. Subjects were excluded from the study if they
153
had one or more of these conditions. Baseline blood pressures and pulse rates were recorded, in
154
addition to height and weight. Twenty-five subjects were studied, with body mass indices of 24.2 ± 3.0.
155
ip t
151
Spectra (540-800 nm) were acquired at a rate of either 0.33 or 0.5 Hz during a 3-min period of rest while each subject breathed room air. An oxygen mask was then placed over the face for 3 min, and
157
the subject breathed 100% O2. A blood pressure cuff was inflated on the upper arm to a pressure that
158
was 60 mmHg above the subject’s systolic blood pressure or to 200 mmHg, whichever was less. The cuff
159
was inflated for 15 min, and the oxygen mask was removed approximately two minutes into the period
160
of arm ischemia. A recovery period of 5 min followed to ensure normalization of blood perfusion to the
161
arm. Spectral acquisition was continuous throughout the entire protocol.
us
an
M
162
cr
156
Consecutive spectra were averaged to improve signal-to-noise ratios. When the acquisition rate was 0.33 Hz, three spectra were averaged together, producing a new rate of 0.11 Hz. When the
164
acquisition rate was 0.5 Hz, five spectra were averaged together for a new rate of 0.1 Hz.
te
d
163
Spectra collected from the hand at rest while breathing 100% O2 represented the most
166
oxygenated condition that we could practically produce in the subjects (Fig.1). We assumed that these
167
spectra represented Mox values of 100% but we have not verified with an independent measure that
168
myoglobin and hemoglobin in the hand were fully saturated with oxygen. Spectra collected at the end of
169
the 15-min period of arm ischemia were assumed to represent Mox values of 0% (Fig. 1). In healthy
170
subjects, it has been observed that 15 min of arm ischemia is sufficient to achieve complete
171
deoxygenation of skeletal muscle and blood in the hand because myoglobin and hemoglobin saturations
172
plateau to a minimum value [12]. Spectra acquired during the period of arm ischemia represent
173
intermediate Mox values between 0 and 100%.
174
2.2. ANALYSIS
Ac ce p
165
Page 8 of 25
5 175 176
All analyses were performed using PLS_Toolbox Version 6.2 (Eigenvector Research, Inc.) running in Matlab R2009b (Mathworks, Inc.). To help remove baseline and background signals due to scattering phenomena, second
178
derivatives with respect to wavelength were taken using the Savitzky-Golay algorithm with a 65-point
179
window and a second order polynomial [16]. The visible wavelength region was restricted to 562.2 –
180
601.7 nm and the NIR region was restricted to 744.7 – 802.2 nm. These wavelength regions were
181
selected because they encompass the absorbances of the oxygenated and deoxygenated species,
182
respectively. The two regions were scaled relative to each other to correct for differences in absorptivity
183
of the species in the visible and infrared regions of the spectra. The magnitude of this adjustment was
184
determined using the method described in section 2.2.2. The two discontinuous wavelength regions
185
were then concatenated. Finally, to correct for changes in illumination and path length, as well as other
186
total-signal influences, each spectrum was normalized to its total summed absorbance.
cr
us
an
M
2.2.1. Estimation of Mox Reference Values by MCR. As no accurate primary method exists to
d
187
ip t
177
determine the actual Mox in human subjects at any given point in time, MCR [17, 18] was used to
189
estimate Mox during ischemia. Separate two-factor MCR models were calculated for each subject to
190
recover estimates of the oxygenated and deoxygenated component spectra and corresponding
191
contributions of each species at each point in time.
Ac ce p
te
188
192
Each model used a closure constraint to assure that the summation of the two recovered
193
contribution profiles would always be 100%. Assuming that the spectra collected just prior to cuff
194
inflation during breathing of pure oxygen are close to 100% Mox, the five spectra at this point in every
195
run were constrained to give a 100%:0% ratio of oxygenated to deoxygenated species. Likewise, the last
196
five spectra in each run prior to cuff release were constrained to give a 0%:100% ratio. The average
197
spectra from each of these time periods were used as the initial guesses for the pure-component
198
spectra in the MCR. Because the closure constraint requires that the two contribution profiles sum to
Page 9 of 25
6 199
100%, the recovered oxygenated species contribution profile is equivalent to the estimated Mox for the
200
subject. An example Mox profile is shown in Fig. 2.
201
It is important to note that this procedure to recover the Mox profiles is only an approximation of the actual Mox, which may be influenced by optical variations (such as skin-probe boundary changes).
203
However, it is also assumed that when all the subjects' results are combined during the LWR modeling
204
later, these effects will be diminished by an averaging of the pool of all subjects.
cr
ip t
202
2.2.2. Correction of Absorptivity Effects. Because there is always an ambiguity in the total
206
volume of muscle interrogated, the total observed absorbance is also ambiguous. The absorbance
207
observed for the deoxygenated species, for example, may increase due to decreased Mox, or due to an
208
increase in total interrogated volume. To correct for these effects, each spectrum is normalized to the
209
total absolute value of the observed signal (the sum of the absolute value) after taking the second
210
derivative and before MCR. This has two side effects: First, it means relative absorbances of the
211
oxygenated and deoxygenated species can be used to calculate Mox. Second, it introduces an
212
unintended nonlinearity between calculated Mox and true Mox due to the difference in absorptivity
213
between the oxygenated and deoxygenated species, as described below. Note that this effect is not a
214
nonlinearity between observed absorbances at different wavelengths (which would lead to a higher
215
multivariate rank of the set of spectra), but one simply between the calculated and true Mox.
217 218
an
M
d
te
Ac ce p
216
us
205
The total absorbance summed across all wavelengths, Aobs, can be written as a summation of the absorbances of the oxygenated and deoxygenated species: Aobs = EoBCo + EdBCd
(1)
219
where B is the effective path length, Eo and Ed are the total absorptivities summed across all
220
wavelengths of the oxygenated and deoxygenated species, respectively, and Co and Cd are the
221
concentrations of oxygenated and deoxygenated species, respectively. If the two absorptivity profiles
222
had the same integrated areas, Aobs would only be influenced by changes in path length or total
Page 10 of 25
7 223
myoglobin + hemoglobin concentration (Co + Cd) and dividing each spectrum by Aobs would correct for
224
changes in those properties. However, because Eo and Ed are not equal, the total summed absorbance,
225
Aobs, will vary with Mox even in the absence of changes in path length or Co + Cd. Fig. 3 illustrates the problem that would occur with normalization if there were a 4:1 ratio
ip t
226
between Eo and Ed. The dotted lines represent "true" oxygenated and deoxygenated species
228
concentrations (in relative %) as Mox decreases linearly from left to right. The solid lines show the
229
concentrations that would be calculated from normalized spectra. At 50% true Mox, the observed Mox
230
would be 80%, reflecting the 4:1 contributions to Aobs from Eo and Ed when Co = Cd. At 100% Mox, the
231
summed absorbance is dominated by the oxygenated species absorbance (Aobs = EoBC), while at 0% Mox,
232
the summed absorbance is dominated by the deoxygenated species (Aobs = EdBC). Aobs at 100% Mox
233
would be four times greater than Aobs at 0% Mox. Thus, as Mox varies, Aobs is not a constant but is
234
changing as a proportion of the absorptivity of the oxygenated or deoxygenated species. The effect of
235
the non-constant Aobs is that observed Mox would be consistently overestimated in this example.
us
an
M
d
One approach to correct for this effect is to scale the visible region of the spectrum (dominated
te
236
cr
227
by the oxygenated species) relative to the NIR region (dominated by the deoxygenated species) by a
238
factor corresponding to the ratio between Eo and Ed. Estimating an appropriate scaling factor was
239
achieved by taking advantage of the fact that when there are no major path length effects (such as in
240
the highly-controlled ischemia trials), normalization has little effect except to introduce the nonlinearity.
241
Once the scaling factor is correct, normalization should have almost no effect on Mox values recovered
242
using MCR. By adjusting the NIR:Visible scaling factor and then examining Mox profiles recovered with
243
and without normalization, we identified a factor at which the profiles were not significantly changed by
244
normalization. This was considered the "optimal scaling factor" for the given subject. The 50% Mox point
245
was used as a reference to indicate how similar the profiles were with and without normalization.
Ac ce p
237
Page 11 of 25
8 246
Figs. 4A and 4B show an example of the change in Mox profile resulting from normalization
247
before a scaling factor was applied. In Figs. 4C and 4D, the 50% Mox points of unnormalized and
248
normalized data sets were identical when a scaling factor of 5.3 was used. This process was repeated for all 20 training subjects and the median optimal scaling factor of
ip t
249
4.7 was selected for all subsequent spectral adjustments. Each training subject's spectra were corrected
251
by the median optimal scaling factor and the Mox profiles were estimated using the MCR procedure
252
described earlier. Finding a single factor to use for all subjects provides for the most general correction
253
of absorptivity and possible optical effects without over-correcting for a single subject. It also provides a
254
means to analyze subjects for which the normalized and non-normalized results were too severely
255
different (probably due to actual path length or optical effects) and enables use of the models for
256
application to future clinical subjects for which we will have only a single point, and not an entire
257
ischemic profile.
us
an
M
2.2.3. Locally Weighted Regression Modeling. Given the estimated Mox values and the
d
258
cr
250
individual spectra for all subjects as a whole, an LWR model was created that used a two principal
260
component PCA model to select training spectra. For any given Mox estimate to be made, 200 training
261
spectra were selected and a one latent variable PLS model was calculated and used for predictions. An
262
attractive feature of such an empirical, inverse least squares PLS model is that having relatively few
263
constraints allows flexibility in correcting for interferences. In contrast, a classical least squares method
264
like MCR will have a greater number of constraints that will likely degrade prediction accuracy.
Ac ce p
265
te
259
The training spectra provided to the LWR model were pre-processed similarly to that described
266
for the MCR estimation procedure (2nd derivative, 4.7 NIR:Vis scaling, area normalization) but the
267
spectra and estimated Mox values were also mean-centered to allow for an intercept in the linear
268
regression. The number of principal components, number of training spectra selected, and number of
269
PLS latent variables to be used were all selected using cross-validation in which the entire set of spectra
Page 12 of 25
9 from a given subject were left out of the training set, and the remaining 19 subjects were used to
271
estimate the left-out subject's Mox. The error in these Mox estimations was summarized in the root
272
mean squared error of cross-validation (RMSECV) which is the square root of the average error of
273
prediction for all samples when they were left out of the training set [14]. This was repeated for various
274
settings of the LWR algorithm with general agreement that the selected LWR parameters gave stable
275
and reliable results.
cr
2.2.4. Partial Least Squares Modeling. For comparison to the LWR models, an identical
us
276
ip t
270
modeling effort was made using standard (non-locally-weighted) PLS modeling. For these models, the
278
identical calibration data were analyzed with a standard PLS algorithm using cross-validation to
279
determine the number of PLS latent variables. The best smallest cross-validation error was observed at 2
280
latent variables.
M
281
an
277
2.2.5. Quality of Spectral Fit Parameters. In both the PLS and LWR models, it is important to assess how well the model fits a given spectrum when making a Mox prediction for that spectrum. For
283
factor-based models such as PLS and LWR, this can be done using the Hotelling’s T2 and Q residuals [14].
284
Generically, these models decompose a data matrix, X (in which rows are spectra), into scores, T, and
285
loadings, P, and spectral residuals, R, according to the equation:
Ac ce p
te
d
282
286
X = TPT + R
287
Hotelling’s T2 is calculated as a weighted sum of the scores and describes whether the features
(2)
288
described by the loadings of the model are present in similar scale to what was observed in the
289
calibration data. The Q residuals are calculated as the sum of the squared residuals in each row of the R
290
matrix [14]. When making a prediction for a new sample, a new Hotelling’s T2 value and Q residual value
291
are calculated and compared to the values obtained for the calibration data. If a new value is
292
substantially different from the calibration data, it is inferred that the new spectrum contains some
Page 13 of 25
10 293
feature(s) which were not observed in the calibration data and, thus, the predictions for Mox should not
294
be used.
295
297
3. RESULTS AND DISCUSSION
ip t
296
3.1. Cross-Validation. While building the models, cross-validation was used as described earlier. For the LWR model, cross-validation indicated very little change as a function of either the number of
299
principal components (PCs) in the model or the number of samples selected for each local model.
300
Models were examined with 1 to 6 PCs, 100, 150, or 200 selected samples, and 1 to 3 latent variables in
301
the local PLS model. The RMSECV ranged from 6.5% at 2 PCs and 200 samples selected for each local
302
model up to 8.1% at 6 PCs and 100 samples selected. In comparison, the best standard PLS model (not
303
LWR-based model) provided an RMSECV of 9.3% obtained with 1 latent variable. All LWR and PLS
304
models with more than 1 latent variable gave RMSECVs above 9.3%. In all cases, the subjects in the
305
calibration set displayed Hotelling’s T2 and Q residuals that indicated there were no unusual subjects or
306
samples in the calibration set.
us
an
M
d
te
3.2. Validation Subject Results. In order to validate the model, Mox measurements were made
Ac ce p
307
cr
298
308
from five healthy subjects that were not included in the model. These subjects underwent the same
309
protocol as the 20 subjects included in the model. Fig. 5 shows that the LWR model produced accurate
310
Mox measurements compared to MCR estimates of Mox obtained from each subject’s data set. The
311
best-fit line had a slope of 0.99 and a y-intercept of -2.2, very similar to the line of unity. The root mean
312
square error (RMSE) for all 654 spectra was 5.4%.
313
For comparison, the results for the standard PLS model applied to the same five test subjects are
314
shown in Fig. 6. Because the PLS model is required to take into account all subjects equally, the model
315
must be a “consensus,” resulting in a per-subject bias. Three of the five subjects show notable
Page 14 of 25
11 316
systematic offset from the MCR estimates. The RMSE for the test subjects using the PLS model was
317
9.6%.
318
Finally, an additional LWR model was built with the wavelength range limited to only the NIR portion of the spectrum (744.7 – 802.2 nm). In this model, the predictions showed little to no trend with
320
Mox and produced an RMSE of 29.3% (Fig. 7). Such a poor result is expected considering that removing
321
the visible portion of the spectrum leaves no ability to correct for path length and absorbance effects. It
322
demonstrates the need for obtaining both the visible and NIR portions together for the Mox evaluation.
323
Interestingly, these poor results are probably overly optimistic considering that reference Mox values
324
were obtained from MCR using the full wavelength range. If the MCR step was similarly limited in
325
wavelength window, no sensible reference values could even be obtained for the calibration samples.
cr
us
an
3.3. Application to Trauma Patients. Mox was measured in a preliminary study of trauma
M
326
ip t
319
patients admitted to Harborview Medical Center in Seattle, WA. Spectra were acquired for 15 min as
328
soon as possible upon admission to the emergency department. The LWR model was applied to the
329
spectra from each patient. Fig. 8 shows the Mox time course in one patient. Mox has a good correlation
330
with blood pressure; it provides a continuous and real-time indication of tissue perfusion. The sharp
331
increases in Mox at approximately t = 2 and 8 min correlate with intravenous saline infusions given for
332
low blood pressure that temporarily improved tissue perfusion. In contrast, arterial blood saturation
333
measured by pulse oximetry was not sensitive to changes in tissue perfusion.
335 336
te
Ac ce p
334
d
327
4. CONCLUSIONS
The results for both the validation subjects and a trauma patient indicate that the use of joined
337
visible and NIR spectra can provide Mox estimates better than what has been previously reported for
338
single- or double-wavelength NIR methods. In addition, these results show that the LWR models are
339
capable of making reasonable estimates of Mox, even when presented with the limited range of
Page 15 of 25
12 340
physiological variability provided in this small calibration set. The accuracy of these predictions should
341
only improve when additional calibration subjects, representing a wider variety of skin tones and body
342
mass indices, are added to the calibration set. Further improvements in model performance would also be expected by using a modified
ip t
343
method for training sample selection. In this work, the sample-to-sample distance in the PCA score
345
space was used. This PCA model captured the variance associated with both the subject-to-subject
346
variability (which is most critical to correct) and the varying Mox caused by ischemia in individual
347
subjects. Thus, samples were selected both on the similarity of the subject as well as the similarity of the
348
extent of ischemia to test spectra. The predictions made from the ordinary PLS model show that the
349
calibration and test subjects were generally in error by a constant slope and/or offset across all Mox
350
values. From this, we would anticipate that a modified sample selection method which focuses more on
351
subject-to-subject variability (for example, discarding all spectra for individuals which are dissimilar and
352
keeping all spectra for individuals which are similar) would tend to create models that capture the
353
specific slope and offset scaling relevant for the tested individual. Such models may provide better
354
predictions than our current model.
us
an
M
d
te
Ac ce p
355
cr
344
This work also demonstrates that the use of MCR with simple equality and closure constraints
356
provides a reasonable inference of Mox in the absence of any true primary reference. This use of a
357
carefully controlled experimental protocol and inferential estimates of Mox can be used to easily
358
produce additional calibration data for the models.
359 360 361
ACKNOWLEDGMENTS This work was funded in part by grants from the Coulter Foundation, the Life Sciences Discovery
362
Fund and NIH 1R21EB009099. None of these granting agencies played any role in the study design or in
363
the collection, analysis or interpretation of the data.
Page 16 of 25
13 364 REFERENCES
366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391
[1] M. Smith, Philos Transact A Math Phys Eng Sci 369 (2011) 4452-4469. [2] J.M. Murkin, M. Arango, Br J Anaesth 103 Suppl 1 (2009) i3-13. [3] T.W. Scheeren, P. Schober, L.A. Schwarte, J Clin Monit Comput 26 (2012) 279-287. [4] B.A. Crookes, S.M. Cohn, S. Bloch, J. Amortegui, R. Manning, P. Li, M.S. Proctor, A. Hallal, L.H. Blackbourne, R. Benjamin, D. Soffer, F. Habib, C.I. Schulman, R. Duncan, K.G. Proctor, J Trauma 58 (2005) 806-813; discussion 813-806. [5] J.C. Hirsch, J.R. Charpie, R.G. Ohye, J.G. Gurney, Semin Thorac Cardiovasc Surg Pediatr Card Surg Annu 13 (2010) 51-54. [6] D. Highton, C. Elwell, M. Smith, Curr Opin Anaesthesiol 23 (2010) 576-581. [7] G.A. Cohen, L.C. Permut, L.S. Arakaki, W.A. Ciesielski, D.M. McMullan, A.R. Parrish, K.A. Schenkman, ASAIO J 57 (2011) 314-317. [8] K.A. Schenkman, D.R. Marble, D.H. Burns, E.O. Feigl, Appl Spectrosc 53 (1999) 332-338. [9] L.S.L. Arakaki, W.A. Ciesielski, B.D. Thackray, E.O. Feigl, K.A. Schenkman, Appl Spectrosc 64 (2010) 973-979. [10] L.S.L. Arakaki, D.H. Burns, M.J. Kushmerick, Appl Spectrosc 61 (2007) 978-985. [11] D.J. Marcinek, W.A. Ciesielski, K.E. Conley, K.A. Schenkman, Am J Physiol Heart Circ Physiol 285 (2003) H1900-H1908. [12] C.E. Amara, D.J. Marcinek, E.G. Shankland, K.A. Schenkman, L.S. Arakaki, K.E. Conley, Methods 46 (2008) 312-318. [13] J.B. Wittenberg, B.A. Wittenberg, J Exp Biol 206 (2003) 2011-2020. [14] H. Martens, T. Naes, Multivariate Calibration, John Wiley & Sons, Chichester, 1989. [15] T. Naes, T. Isaksson, B. Kowalski, Anal Chem 62 (1990) 664-673. [16] A. Savitzky, M.J.E. Golay, Anal Chem 36 (1964) 1627. [17] M.H. Van Benthem, M.R. Keenan, D.M. Haaland, J Chemom 16 (2002) 613-622. [18] P.J. Gemperline, E. Cash, Anal Chem 75 (2003) 4236-4243.
cr
us
an
M
d
te
Ac ce p
392
ip t
365
Page 17 of 25
14
us
cr
ip t
392
an
393
Figure 1. Spectra acquired from a healthy human subject at extremes of muscle oxygenation. The solid
395
spectrum was acquired after 3 min of 100% O2 breathing. The dashed spectrum was acquired after 15
396
min of arm ischemia. Oxymyoglobin and oxyhemoglobin absorb near 580 nm in the visible region;
397
deoxymyoglobin and deoxyhemoglobin absorb near 760 nm in the near infrared region. The visible and
398
NIR bands that were used to estimate Mox are indicated by the sets of vertical dashed lines. In the
399
wavelength region between 610 and 730 nm, the detector was saturated. These spectral values do not
400
represent measurements of tissue absorbance and are not used in the spectral analysis.
d
te
Ac ce p
401
M
394
Page 18 of 25
15 401
an
us
cr
ip t
402
403
Figure 2. Muscle oxygenation (Mox) in a healthy human subject estimated by MCR. The subject
405
breathed air for 3 min, then 100% O2 for 3 min, and then underwent a 15-min period of arm ischemia.
M
404
Ac ce p
te
d
406
Page 19 of 25
16 406
an
us
cr
ip t
407
408
Figure 3. Normalization bias effect. Since the integrated areas of absorptivity profiles of oxygenated and
410
deoxygenated species are not equal, normalization introduces an unintended nonlinearity to recovered
411
relative oxygenated and deoxygenated concentrations. Dotted lines represent true relative
412
concentrations and by definition, true Mox is equal to “True Oxygenated.” True Mox decreases from left
413
to right, as during ischemia. If the ratio of the integrated areas of oxygenated to deoxygenated
414
absorptivity profiles were 4:1, the observed relative concentrations would be those shown by the solid
415
lines. The observed Mox, equal to “Observed Oxygenated,” would be consistently overestimated.
d
te
Ac ce p
416
M
409
Page 20 of 25
17 416 417 418
ip t
419 420
cr
421
Ac ce p
te
d
M
an
us
422 423 424 425 426
427 428 429 430 431 432 433 434 435 436 437 438 439 440
Page 21 of 25
18 Figure 4. Scaling correction for absorptivity effects. A) MCR results without (left) and with (right)
442
normalization show that the 50% Mox point is affected by normalization. B) Concatenated second-
443
derivative spectra. The visible region comprises 562-601 nm and the NIR region comprises 744-802 nm.
444
C) Up-weighting the NIR region where deoxygenated species have predominant intensity corrects for
445
the normalization effect. This example shows results when the NIR region was up-scaled by a factor of
446
5.3. The optimal scaling factor was determined by adjusting the scaling until the 50% Mox points for
447
unnormalized (left) and normalized (right) cases were equal. D) Concatenated second-derivative spectra
448
with the scaling, without and with normalization.
us
cr
ip t
441
Ac ce p
te
d
M
an
449
Page 22 of 25
19
us
cr
ip t
449
an
450
Figure 5. Accuracy of muscle oxygenation (Mox) measurements by LWR in five healthy human subjects.
452
Mox was measured accurately across its range from 0% to 100%.
M
451
Ac ce p
455
te
454
d
453
456 457 458 459 460 461 462 463
464
Figure 6. Accuracy of muscle oxygenation (Mox) measurements by PLS in five healthy human subjects.
465
Most subjects show a systematic bias at all Mox levels.
Page 23 of 25
us
cr
ip t
20
466
Figure 7. Muscle oxygenation (Mox) measurements by LWR in five healthy human subjects when only
468
the NIR portion of the spectra is used. Accuracy is very poor. The advantage of including the visible
469
region in the analysis is apparent when these results are compared to those in Figure 5.
M
an
467
470
d
471
Ac ce p
te
472 473 474 475 476 477 478 479
480
Figure 8. Muscle oxygenation (Mox) is indicative of tissue perfusion. Mox was measured in a trauma
481
patient who received saline infusions at t = 2 and 8 min (indicated by arrows), causing increases in blood
482
pressure (BP) and corresponding increases in Mox. Arterial saturation measured by pulse oximetry was
483
not sensitive to changes in tissue perfusion.
Page 24 of 25
ce Ac
Page 25 of 25