WO1999043255A1

WO1999043255A1 - Near infrared-transmission spectroscopy of tongue tissue

Info

Publication number: WO1999043255A1
Application number: PCT/US1999/004054
Authority: WO
Inventors: Gary W. Small; Jason J. Burmeister; Mark A. Arnold
Original assignee: University Of Iowa Research Foundation; Ohio University
Priority date: 1998-02-25
Filing date: 1999-02-25
Publication date: 1999-09-02
Also published as: AU3311099A

Abstract

Non-invasive measurement of blood glucose levels is carried out using transmission spectroscopy on the tongue of the individual. The near infrared (NIR) first overtone spectral region between 7000 and 5000 cm-1 (1.43 - 2 νm) is well suited for non-invasive measurements. Because the tongue is one of the leanest accessible sights in the body, variations in readings due to body fat are minimal. Further, the tongue is well thermostated and vasculated which are important factors in selecting a measurement site.

Description

WO 99/43255 . \ . PCT/US99/04054

NEAR INFRARED-TRANSMISSION SPECTROSCOPY OF TONGUE TISSUE

DESCRIPTION Background of the Invention

This application relates to the non-invasive measurement and monitoring of chemical substances found in blood, and particularly to the non-invasive measurement and monitoring of blood glucose levels.

Diabetes affects approximately 16 million people in the United States. People suffering from diabetes mellitus must monitor their blood glucose levels for insulin replacement therapy. At present, this generally requires an invasive procedure, performed several times daily, in which a sample of blood is taken for analysis. Non-invasive procedures have been proposed, including those found in US Patents Nos. 5,069,229 and 5,459,317, and EP 0 548 418, which are incorporated herein by reference, but no in-home non-invasive glucose measurement device is as yet available.

Among the proposals for non-invasive measurement of blood chemicals are those which depend on diffuse-reflectance of infrared radiation, for example from a fmger, lip or forearm. In these techniques, light is delivered to the surface of the tissue, penetrates for some distance and then is reflected back to be detected. The drawback of the diffuse reflectance measurements is that the depth to which the light penetrates is not deep enough to have enough interaction with glucose to provide consistent and quantitative detection of glucose. In addition, tissue areas conventionally used for diffuse reflectance measurements may have varying amounts of fat from individual to individual. Since fat interferes with detection of glucose, this variability can make calibration and quantification difficult.

It is an object of the present invention to provide an improved method for the non-invasive measurement of glucose.

Summary of The Invention

The present invention overcomes the difficulties of prior art processes by using transmission spectroscopy, and by performing the measurements on the tongue of the - 2 -

individual. The near infra red (NIR) first overtone spectral region between 7000 and 5000 cm^"1 (1.43 - 2 μm) is well suited for non-invasive measurements. Because the tongue is one of the leanest accessible sights in the body, variations in readings due to body fat are minimal. Further, the tongue is well thermostated and vasculated which are important factors in selecting^' a measurement site.

Brief Description of the Drawings

Fig. 1 shows SEP versus noise plots for model data sets. Vertical lines indicate noise levels of the webbing (solid) and tongue (dashed) for back to back spectra at different measurement wavelengths.

Figs. 2A-F show absorbance spectra for near-infrared measurements at different measurement sites.

Fig. 3 shows single beam spectra of human webbing (dashed) and tongue (solid) with a glucose absorbance spectrum (dotted) overlaid.

Figs. 4 A and B show 100% lines of back-to-back psectra of human webbing and tongue with an absorbance spectrum of 1M glucose overlaid.

Figs. 5 A-E show glucose concentrations versus sample number for the calibration samples for five individuals.

Figs. 6 A-D show SEP versus RMA noise calculated over 100 cm^"1 spectral ranges for model data sets. Circles and triangles are 5.6 and 6.3 mm aqueous thickness without scatter, respectively. Squares are prediction errors from the model data set with scattering particles added. The vertical lines are noise levels of subject D who had an aqueous layer thickness of 5.5 mm.

Figs. 7A-E are concentration correlation plots for the best PLS models for the five volunteers predicting the blind samples.

Figs. 8A-E are concentration plots for the five volunteers, using the first sample as the calibration and the last samples as the prediction. - 3 -

Detailed Description of the Invention

In accordance with the present invention, non-invasive monitoring of glucose in a human patient, is carried out by measuring the near infrared absorbance of glucose using transmission spectroscopy on the tongue of the human patient. This selection of measurement site provides better correlation with invasive measurements than non-invasive measurements performed at other sites, including the index fmger and webbing of the hand.

In selecting a site for non-invasive glucose determination, several factors need to be considered. First, a useful technique needs to produce results which are reliable predictors of the actual glucose levels. Second, it would be desirable to have a calibration model which would work in substantially all patients. Patient-to-patient variation in the physical size of the tissue being measured and in the levels of interferents such as fats, which can act as a major interferent in the measurement of glucose in the overtone region, can make achieving these goals a complicated process. Our evaluation of several alternative measurement sites, however, indicates that the tongue is a highly suitable location for non- invasive glucose measurements.

To evaluate various sites for their suitability, single beam transmission spectra were collected of human webbing between the thumb and forefinger, tongue, nasal septum, cheek, upper lip and lower lip. As described below, the hand webbing had the highest levels of fat in the individuals tested, thus producing the worst results, while the tongue had the lowest fat levels and produced the most accurate and reproducible results.

- 4 -

EXAMPLE 1 Evaluation of Alternative Measurement Sites Spectra were collected on a modified Midac spectrometer for the near infrared range. Modification include laser detectors, a fan designed to control temperature in the spectrometer, a 250-Watt source operating at about 110 watts and a one-millimeter diameter InGaAs detector with a 1.9 μm cutoff (Epitaxx). An H-band astronomical filter was used to limit the source bandpass to utilize the full dynamic range of the detector without saturation and minimize sample heating. A one-inch diameter, 25 mm focal length convex lens was used to focus the source intensity on the sample or fiber optic bundle.

Preliminary single beam spectra were collected using a fiber optic bundle olan

Jenner Industries) connected to the thickness controlling window to direct the incident

radiation to the sampling site. The detector and mount was freed from the spectrometer

for ease of sampling. This arrangement allowed spectra to be collected from hard to

reach body sites. Single beam transmission spectra were collected of human webbing

between the thumb and forefinger, tongue, nasal septum, cheek, upper lip, and lower lip.

A background single beam spectrum was collected without a sample in the optical path

using a combination of a neutral density filter to limit the incident source intensity and a

low detector preamplifier gain setting to avoid detector and electronic saturation. The air

spectrum was corrected to match the higher gain and incident radiation intensity of the

tissue spectra before converting it to absorbance.

Spectra with lower noise levels were collected without the fiber optic bundle to

maximize incident source intensity. The detector and mount were fixed in the spectrometer in a vertical position for the tongue spectra and in a horizontal position for

the webbing spectra. A mirror was used to direct the incident radiation to the detector for

the tongue spectra. - 5 -

Absorption spectra of the webbing of 19 volunteers were collected. A regression

analysis was performed to compute the thickness of water and fat from the spectra using

beef fat and water as the standards. The average thickness of water was 4.67 mm and

beef fat was 2.4 mm for a webbing thickness of 6 mm. Table 1 lists RMS noise levels

for various frequency ranges of ten back-to-back webbing spectra of 128 scans each.

Figure 1 shows plots of SEP versus RMS noise levels for various frequency ranges for

model spectra consisting of a constant fat thickness of 1.6 mm and variable water thicknesses of 5.6 and 6.2 mm. From Figure 1 and Table 1 , a very high SEP would

be expected if the webbing were used (greater than 3 mM). The RMS noise levels are

low in the spectral region which is dominated by water absorbance (>6000 cm^"1). The

noise levels in the region dominated by fat absorbance are very high. They are an order

of magnitude higher than the model spectra. Between measurement variations in the amount of fat compressed within the light path have not been successfully controlled.

Another body site for the noninvasive measurement is needed because of the strong

overlap between the absorbance bands of fat and glucose.

Table 1 RMS noise values calculated from webbing and tongue 100% lines.

Spectral Range (cm^"1) 6400 6300 6200 6100 6000 5900

Sample -6300 -6200 -6100 -6000 -5900 -5800

6 mm webbing 54.7 25.1 18.0 36.1 271.1 605.8 5.5 mm tongue 96.7 47.7 25.4 18.0 56.6 68.1 - 6

The composition and physical thickness of several bodily sites were evaluated with the goal of finding a suitable site for a noninvasivc measurement of glucose. Absorption spectra are listed in Figures 2A through 2F. It is apparent from the

spectra that the upper and lower lips, tongue, cheek and nasal septum have less fat than

the webbing. A regression analysis was performed to quantify the amount of fat and

aqueous thickness. Table 2 lists the calculated aqueous and fat thicknesses. The

tongue was found to be most desirable because it has the least amount of fat and there are

relatively few variations between people. Major disadvantages of the tongue are the

possibility of spreading disease while sampling many people and the discomfort of

collecting spectra.

Table 2 Calculated aqueous and fat thicknesses obtained from regression analysis.

Sample Thickness Aqueous thickness Fat thickness Scatter (mm) (mm) (mm)

Webbing 5.75 4.91 2.62 1.56

Upper lip 6 5.91 0.48 1.72

Lower lip 6 5.95 0.65 1.84

Tongue 6 5.61 0.26 1.47

Nasal septum 6 5.86 0.51 1.81

Check 6 5.61 0.37 1.61

- 7 -

Spectra were collected through the tongue of various people. The average thickness of fat and water was 0.2 and 5.9 mm, respectively, for the 10 volunteers who

were sampled. Figure 3 shows single beam spectra of the human tongue and webbing

collected under the same instrumental configuration with a superimposed absorbance

spectrum of 1.0 M glucose. Although the spectrometer alignment was not held constant,

the spectrum of the tongue is less absorbing and has less fat. Figure 4 shows 100%

lines of back-to-back spectra of the human webbing and also from human tongue. Large

variations in the fat bands which overlap with the glucose bands at 5920 and 5750 cm^' in

the webbing spectra makes it obvious that the tongue is a superior measurement site. Table 1 also shows noise levels of tongue spectra calculated in the hospital with

multiple interfaces. From these noise levels and .Figure 2 one may estimate a SEP of

about 2 mM on a one subject study.

It is clear that the tongue is a superior measurement site for a noninvasive

measurement compared to the webbing. The convenience of collecting spectra through

the webbing of the hand is traded for lower variations in the amount of fat within the light

path. The interface between the spectrometer and the volunteer will need to be sterilized

between collecting spectra from different subjects. If the same spectral quality were

obtained in a noninvasive data collection as the test spectra from the tongue a SEP of about 2 could be obtained. - 8 -

EXAMPLE 2 Detection of Blood Glucose Levels Using Transmission Spectroscopy of The Tongue

Spectra were collected over the course of 39 days on five type 1 diabetic

volunteers, 3 female and 2 male. The experimental procedure was approved by a human

subject protocol board at the University of Iowa Hospitals and Clinics. Two spectra and

one reference blood glucose concentration were collected for each volunteer's visit which

numbered greater than 100. The number of visits per day were limited to less than five

per volunteer.

Spectra were collected using a Midac M-Series spectrometer modified for the

NIR. The spectrometer modifications are detailed in Chapter 8. The source was held at

81.0 watts by supplying constant voltage. An H-band astronomical filter was used to

limit the bandpass of the source intensity. The 1-mm diameter 1.9 μm cutoff InGaAs

detector (Epitaxx) was held at 15 °C by supplying constant voltage to a thermal electric

cooler. The resistor in the feedback loop of the first stage of the detector preamplifier

was 27.1 kΩ. Interferograms (8K) were collected, Fourier transformed, and the resulting

single beam spectra with a data spacing of 4 cm^"1 were saved over the spectral range

between 7000 and 5000 cm^"1. The temperature of the interface was held at 38.5 °C using

a DC heating element and a temperature control circuit. The temperature was monitored

at the following cites during the course of each data collection: room, interface, and

inside the spectrometer. - 9 -

Five interchangeable stainless steel interfaces were machined. These consisted of

a cylindrical sheath to cover the detector and an adjustable lever which controls the sample thickness. Each piece had a hole machined through with a 1 1.5-mm diameter

circular sapphire window to allow the source radiation to pass through the sample to be

detected. The thickness of the sample was confined to 5.45 mm. The interchangeable

interfaces were swabbed with 95% ethanol and autoclaved at 130 °C for 5 minutes

between volunteer visits.

Reference Blood Glucose Values

Capillary blood from a finger stick was drawn into a 100 μL volume heparinized

tube. These samples were analyzed immediately with a YSI 2300 Stat Plus glucose

analyzer (Yellow Springs Instruments) with both probes measuring glucose. These

reference values were validated with a simultaneous measurement using a One-touch test

strip capillary blood glucose analyzer. The average of two YSI glucose values was used

to build PLS models. Every fifth blood sample was considered a 'blind' and was not used

to optimize the PLS models.

Data Analysis

Separate PLS models were optimized using single beam spectra from each

individual. The first 189 spectra were not included because the source voltage varied and

the volunteers were learning how to perform the measurement in this time period.

Spectra were also removed if their single beam intensity was very low. A PCA score analysis was performed but no spectral outliers were removed. The PLS calibration - 10 -

model parameters were optimized. Three unique rearrangements of calibration and monitoring sets were used to optimize spectral range and number of PLS factors. The parameters which gave the lowest average SEM for all three rearrangements of calibration and monitoring were selected. This procedure reduces the chance of over modeling and modeling information specific to a single small monitoring set .

Concentration Correlations

By limiting the number of samples that were collected each day, the correlation

between glucose and time was minimized. Figure 5 shows plots of glucose versus

sample number for the calibration samples for the five subjects. They look like scatter

plots. Table 3 shows the correlation coefficient (r²) of the time profiles. The

correlation coefficients are all below 0.014. It is not likely that any temperature variation

or spectrometer drift correlated with these randomized concentrations. In this way,

spectral variations that the PLS models are built upon should only be analyte specif tiicc changes.

Table 3 Description of glucose concentrations.

subject number high low mean standard deviation Regression coefficcnts of time plots (mM) (mM) (mM) (mM) Bo B, R²

A 173 21.3 2.94 9.91 4.53 8.904 0.01082 0.01377

B 178 28.3 2.54 12.64 6.67 12.13 0.006442 2.465c-3

C 181 19.1 2.83 9.05 4.06 12.62 -2.203e-4 6.085c-6

D 207 26.85 3.20 1 1.88 5.09 8.625 5.032c-3 4.245c-3

E 169 16.75 3.79 8.57 2.83 8.793 -8.642c-4 2.020c-4

- 11 -

Significant glucose concentration variation is seen in all subjects except volunteer

E who has very tight control over her glycemia. Table 3 ij_sts the high, low, mean, and

standard deviation of each subject's sample concentrations as well as the total number of

spectra. The subjects were not persuaded to vary their daily diabetes treatments. The

glucose concentrations are believed to be representative of the blood glucose variation for these individuals. Significant spectral and concentration changes are needed to build

meaningful PLS calibration models. The limited concentration range for volunteer E

might make calibration difficult.

Spectral Quality The spectral quality is fairly high. Table 4 lists RMS noise levels calculated

over 100 cm^"1 intervals, the standard deviation of the ratio of the intensity at 5751 to 6994 cm^' , standard deviation of the single beam intensity, and signal to noise ratio. The RMS

noise levels were calculated using a second order polynomial fit for 100% lines from

back-to-back single beam spectra. The noise levels between 6100 and 6000 cm^'1 are

fairly low but the noise levels between 6000 and 5900 cm^"1 are not as low as those

collected on the author in the pre-experiment trial. The ratio of the intensity at 5751 to 6994 cm^'1 is the intensity at the fat band divided by the intensity at the peak of the single

beam, respectively. The signal to noise (SNR) is the peak single beam intensity divided

by the RMS noise of the single beam calculated between 7000 and 6800 cm^'1. The

SD/mean intensity is the standard deviation of the peak single beam intensities divided by

the mean peak intensity multiplied by 100%. - 12 -

Comparison to Model

The in vitro model was used to estimate ideal prediction errors for the noninvasive data set. The aqueous layer thickness was estimated to be about 5.5 mm. By comparing

the noninvasive noise levels to Figure 6 _; jt was estimated that an SEP of about 2.3 M

would be the best that could be expected for the noninvasive data if there were no

interferences and no variation in the amount of fat in the optical path. The vertical lines

are the noise levels from subject D. The SEP estimate for the 6000-5900 cm^"1 plot is

found by drawing a line between the models with and without scattering particles.

Table 4 Description of noninvasive spectral quality.

Subject 6100-6000 6000-5900 SD Intensity SD/mean SNR

(μA.U.) (μA.U.) (5751/5994 cm^"1) intcnsity(%) (xlO⁴)

A 26.9 95.1 0.007191 7.71 4.2 B 28.0 88.3 0.009485 8.08 3.5 C 26.6 89.8 0.005072 8.34 4.2 D 24.5 82.0 0.006365 9.64 3.9 E

31.1 99.1 0.011918 9.62 3.7 13

Predicting Blind Samples The calibration data set was used to predict the blind samples using the optimal

PLS parameters. Figure 7 A-E are concentration correlation plots for the best

PLS models for the volunteers predicting the blind samples. The circled points are

spectra in the blind that are not within the scatter of the calibration data. The circled

point for volunteer D at about 1 1 mM was had the highest normalized intensity at 5943 cm^"1 which is where protein absorbs. The circled point for volunteer C at 19 mM is the

highest concentration sample in the data set. There seems to be a bias of under predicting

high concentration samples. This bias was also observed in monitoring sets. The best SEP's are 2.70 and 2.66 for volunteers D and C with the high concentration samples

removed, respectively.

Table 5 Optimal PLS parameters and prediction errors from predicting blind samples.

Who ^■ Range Factors SEC SEP. SEP_b SDP (oV) (mM) (mM) (mM) (mM)

A 6510-5600 1 1 3.51 3.48 4.41

B 6310-5550 1 1 4.29 7.56* 5.50 6.75

C 6600-5650 1 1 3.04 3.49 2.66 3.73

D 6550-5620 16 2.52 3.96 2.70 4.69

E 6600-5650 10 2.62 3.35* 2.74

SEP_a includes all data in blind

SEPb excludes circled data * SEP is greater than SDC - 14 -

In Figure 7B for subject B one point is not shown. The model predicted a blind

sample at 48.9 M when the actual concentration was 16.5 mM. The SEP without this

point is 5.50 mM. The model for subject E is predicting the mean concentration., The

slope is actually negative for this concentration correlation plot (Figure >7E). Even though the model has no prediction, 88% of the prediction points are in the A and B

regions due to the limited concentration range of this subject. ' Table 6 lists the SEC,

SEP, regression analysis and percentage of predicted samples in each of the regions on the Clarke error grid.

Table 6 Prediction errors and Clarke error grid analysis of predicting blind samples.

subject SEC SEP SDP Regression Analysis % of prediction in Clarke error grid Regions

( M) ( M) (mM) ol^" Prediction Data

Bo B, R² A B C D E

A 3.51 3.48 4.41 4.031 0.5466 0.4333 54.1 37.8 8.1 0 0

B 4.29 7.56* 6.75 10.01 0.3393 0.1 198 28.9 52.6 13.2 5.3 0

C 3.04 3.49 3.73 5.423 0.3710 0.2274 43.2 51.4 0 5.4 0

D 2.52 3.96 4.69 5.992 0.4543 0.3136 53.7 34.1 4.9 7.3 0

E 2.62 3.35* 2.74 8.778 -0.05689 0.0156744.1

44.1 0 11.8 0

- 15 -

Predicting Last Part of Data In a real world situation for in-home or clinical devices there will be an initial data

collection used to build the calibration set. The model must be robust so it can accurately

predict the concentrations of spectra following the initial calibration. Predicting the blind

samples which were every fifth sample throughout the entire data collection is unrealistic

in that there is a calibration sample before and after each blind sample. Spectrometer or

temperature drifts could be more easily accounted for in this situation. To approximate a

real world calibration, spectra from the beginning of the data collection were used as the

calibration and spectra from the end of the data collection were predicted. The same PLS parameters used to predict the blind samples were used. Figures 8A through E _are

concentration correlation plots of the five volunteers using the first samples as the

calibration and the last samples as the prediction. Table 7 shows details of the data

sets and model performance.

Table 7 Prediction errors and Clarke error grid analysis of predicting last part of data.

subject SEC SEP SDP Regression Analysis of % of prediction in Clarke error grid Regions

Prediction Data (mM) (mM) (mM) B₀ B, R² A B C D E

A 3.27 4.42 4.72 7.562 0.2623 0.1643 26.9 55.8 5.8 11.5 0

B 4.77 6.08 6.58 10.69 0.2379 0.3129 17.6 52.9 20.6 2.9 5.9

C 2.96 3.86 4.1 1 8.027 0.3358 0.3067 40.7 50 5.6 1.9 1.9

D 2.73 3.35 5.44 4.1 19 0.6930 0.6303 54.2 35.4 2.1 10.4 0

E 2.71 2.72 2.66 8.577 0.06652 0.02379 43.8

5 566..22 0 0 0 - 16 -

In the noninvasive spectra there arc two sources of RMS noise on 100% lines.

The first is spectral noise from the detector and other electronic and optical components

of the spectrometer. This noise is related to the number of photons at a given frequency

which are detected by the detector. The in vitro model of a aqueous and fat layer of constant thickness produces this type of noise. The other type of noise is variation of strongly absorbing components such as fat and muscle. Although the noninvasive spectra

have low noise in the spectral region dominated by water absorbance (greater than 6000 cm^"1), there are variations in the fat and muscle bands (below 6000 cm^"1 ) which are

believed to be causing our high prediction errors. Subject E who had the worst PLS

model had the highest variation in the fat band. This variation could be related to body

composition or movement during the data collection.

The concentration correlation plot for subject E illustrates a situation where all

predictions are in the A and B regions on the Clarke error grid but the model is clearly

void of any analytical utility. This shows that the Clarke error grid alone is insufficient to judge analytical performance of PLS models.

Although, it is debatable whether glucose is being modeled for subjects A, B, and

C. There is very strong evidence that glucose is being modeled for subject D. There is a

bias of under predicting high concentration spectra. If this were due to a lack of

analytical information the lower concentrations would be over predicted as with subject

E. This is not the case. The predictions at low concentrations actually follow the unity

line. It has been reported that there is a time delay in the increase of interstitial glucose

concentration when blood glucose concentration is rising. When light passes through - 17 -

thc tongue it is believed that the majority of the photons will interact with interstitial fluid

not blood. The mismatch between our predicted glucose values and reference blood

glucose values at high concentrations accounts for scatter at high concentrations. There arc conflicting claims in the literature about the exact relationship between blood and interstitial glucose concentrations. If the lower interstitial glucose docs not track blood

glucose at high concentrations, it would be further evidence that we are actually

measuring glucose. This has never been observed before and is further evidence that we

are measuring glucose absorbance information.

For the first time, a human noninvasive blood glucose measurement has been demonstrated. The calibration was uch that glucose concentration was independent of sample order which eliminates the possibility of falsely low prediction errors due to

chance correlations between glucose concentration and temperature or spectrometer drift.

A real world calibration where data collected initially was used to predict data collected at

a later time yielded similar prediction errors as predicting the blind samples which were

every fifth sample throughout the data set.

For future measurements, an improved design for the interface may improve

spectral reproducibiiity, thereby, reducing prediction errors. Also, the source power could

be increased by at least 50% which would yield a higher signal to noise ratio. Although,

spectral quality was probably compromised, the two-month data collection better approximates real world situations compared to calibrations over the course of one week.

Claims

1. A method for non-invasive monitoring of glucose in a human patient, comprising measuring the near infrared absorbance of glucose using transmission spectroscopy, wherein the measurement is performed on the tongue of the human patient.

2. The method of claim 1 , wherein the near infrared absrobance is measured in the first overtone spectral region between 7000 and 5000 cm^"1.