Glucose monitoring using implanted fluorescent microspheres

Working Toward A Minimally Invasive Means for Diabeticsto Better Monitor Glucose Levels

Diabetes mellitus is a metabolic disorderin which insulin, the hormone which

promotes the uptake of glucose into cells,cannot properly perform its role. Becauseglucose is the body’s main energy sourceand is the only carbohydrate metabolizedby the brain, retina, and germinal epithe-lium of the gonads, the disease has a pro-found effect on physiological function[1]. It is estimated that 16 million peoplein the United States and over 100 millionpeople worldwide suffer from this debili-tating disease and its complications.Based on 1996 death certificate data, dia-betes was the seventh-leading listed causeof death in the US, contributing in193,140 cases [2].

The Diabetes Control and Complica-tions Trial (DCCT), published in 1993 byThe National Institute for Diabetes and Di-gestive and Kidney Disorders (NIDDK),indicates that the onset of secondary ef-fects can be delayed and the severity can beminimized with proper care [3]. The care-ful monitoring and maintenance of bloodglucose concentration, according to thestudy, involves injection of insulin or di-etary adjustment after a minimum of fourglucose measurements per day. Despite thetremendous incentive of reducing the de-bilitating complications, very few patientscomply with the recommendations due tothe discomfort associated with the currentpainful, invasive methods for blood with-drawal. Therefore, it is desirable to providediabetics with a simple and painlessmethod for monitoring blood glucose lev-els, which would allow them to better con-trol their conditions.

In this article we discuss our contribu-tion to solving this problem: the uniqueconcept of transdermally queriable,implantable carrier microspheres, encap-sulating a fluorescence assay. Given thesuccess of energy-transfer-based glucose

assays in solution [4-7] and immobilizedwithin membranes at the tip of optical fi-bers [8] , the development ofbiocompatible beads incorporating asimilar biochemical assay is a realisticgoal. In such a system, glucose-perme-able hydrogel microspheres implantedwithin the dermis would respond to theglucose levels in the interstitial fluid.Then, fluorescent light, with characteris-tics dependent upon glucose concentra-tion, would be sensed through the tissueafter excitation from an external or inter-nal source. Because of the positive pre-liminary results achieved in glucosesensing with fluorescence spectroscopy,and the ability to produce biocompatiblematerials for assay encapsulation, theconcept of implantable glucose-sensitivemicrospheres is feasible.

Sensor/InstrumentationGroundwork

The dextran/concanavalin A (Con A)competitive binding system has beenstudied previously and has shown prom-ise for glucose measurement in solution[4, 8, 9]. Therefore, this combination waschosen as a starting point for encapsula-tion studies. The basis of this system as afluorescence assay is the phenomenon offluorescence resonance energy transfer(FRET), where energy is transferrednonradiatively from an excited moleculeto a different molecule in close proximity[10]. Until recently, the FRET glucose as-says had been used only in solution [4-7]and as immobilized within a membrane atthe end of optical fibers [8-9], though ref-erences have been made to implantationand the possibility of using polymericprotection for the assay chemistry [7,11].

The challenges in developing animplantable sensor are to internally con-fine the glucose assay chemistry, to allowrapid diffusion of glucose, and to establisha host-friendly interface at the tissue/fluidcontacting surfaces. Advances in polymerscience have brought the use of hydrogelssuch as poly(ethylene glycol) (PEG) tomany biomedical applications, as they donot evoke significant immune responsewhen implanted. For example,PEG-based polymers have been tested asdrug-delivery devices and biocompatiblemembranes for electrochemical sensors[12,13]. It has also been shown that theydo not limit glucose transfer [12].

Within the past two years, theFITC-dextran/TRITC-Con A glucose-specific assay has been encapsulated inhydrogel spheres [14,15]. First, alginate-poly-L-lysine was used in an attemptphysically immobilize the sensor chem-istry, but the resulting spheres exhibitedpoor structural stability over time and ex-

36 IEEE ENGINEERING IN MEDICINE AND BIOLOGY November/December 20000739-5175/00/$10.00©2000IEEE

Michael J. McShane,*1 Ryan J. Russell,2

Michael V. Pishko,2 and Gerard L. Coté1

*Biomedical Engineering ProgramLouisiana Tech University

1Biomedical Engineering ProgramTexas A&M University

2Chemical Engineering DepartmentTexas A&M University

©19

97Ph

otoD

isc,I

nc.

November/December 2000 IEEE ENGINEERING IN MEDICINE AND BIOLOGY 37

perienced leakage of the biomolecules[14]. While the resulting sensors were re-sponsive to glucose increases up to 600mg/dL when tested in short-term studies,reversal of the glucose binding was notachieved within a reasonable time. Thesame assay was also encapsulated inPEG spheres, where TRITC-Con A waschemically immobilized through linkageto α-acrylol, ω-N-hydroxysuccinimidylester of poly(ethylene glycol)-propionicacid (PEG NHS) prior to polymerization,and FITC-dextran was physicallytrapped by the gel matrix [15]. Thecrosslinked spheres showed homoge-nous distribution of the encapsulatedsensor chemistry and were stable againstfracture and leaching. These spheres, inthe best case, showed a linear response toincreasing glucose beyond 1000 mg/dL,with response times of approximately 10min for a change from 0 to 200 mg/dL.While the results suggest that PEG-basedspheres may eventually be developed forimplantable glucose sensors, a true re-versible response with the dextran/Con Asystem has not yet been realized.

At the same time the details of the sen-sors are being addressed, parallel effortsproceed in development of appropriate in-strumentation and calibration algorithms.Given the problem of designing a probefor measurements of sensors implanted inthe skin, a Monte Carlo simulation wasused to describe the system in terms ofphoton propagation and to aid in the de-sign of a fiber-optic probe for in vivo use[16]. Additional design criteria for the op-tical system were portability, probe flexi-bility, sensitivity, versatility, andrejection or removal of the excitationspectrum. The probe system to meet theseneeds was developed and previously de-scribed [17,18]. This article focuses onexperiments performed using this systemto evaluate the potential for trans-cutaneous monitoring of implanted fluo-rescent particles by investigating tissueoptics and calibration.

Materials and Methods

InstrumentationThe optical probe was assembled using

seven fibers with 400 µm core/425 µmcladding diameters, suitable for visible andnear-infrared transmission (Thorlabs, Inc.,Newton, NJ). A six-around-one configura-tion was used to minimize dead space be-tween fibers. The probe was fixed flush tothe front face of a methacrylate cuvette us-

ing a custom-built sample chamber with anSMA connector, allowing efficient excita-tion and collection of samples. The outputof a 200 W Hg(Xe) arc lamp (Oriel,Stratford, CT) was collimated andapertured, then directed into a 43X micro-scope objective mounted in a three-dimen-sional (3-D) linear and two-dimensional(2-D) tilt stage. The focused output lightwas centered on the excitation fiber, pro-viding 19.5 µW at the distal probe head.

The six collection fibers were coupledto a SPEX 270M (Instruments SA, Edi-son, NJ) imaging spectrograph using astage designed for multiple-fiber inputsand was manufactured in-house. Thechuck was built to interface with the inputport of the spectrograph and contained agrooved plate for positioning fibers. Lightentering the spectrograph passed througha 5 cm-diameter, 500 nm longpass filter(Oriel, Stratford, CT) and was reflectedby mirrors to a 300 grooves/mm diffrac-tion grating blazed at 450 nm. The dis-persed light was then directed to a CCDcamera (Princeton Instruments, Prince-ton, NJ) mounted at the exit port of thespectrometer. The grating was positionedat a center wavelength of 560 nm, suchthat the spectral range detected by the de-tector array was approximately 500 to 620nm. This coverage is sufficient for the ma-jority of the emission spectra produced bythe fluorophores considered in this work(FITC and TRITC). A diagram of thecomplete optical system is found in Fig. 1.

Reagents and Tissue SamplesDextran (MW 2,000,000), FITC,

TRITC, TRITC-Con A, FITC-dextran, andTRIZMA acid and base were purchasedfrom Sigma Chemical Co. (St. Louis, MO).Glucose was purchased from EM Scientific(Gibbstown, NJ) in crystalline form. Twosets of fluorescent microspheres (15 µm di-

ameter) with yellow-green (515 nm emis-sion) and orange (560 nm emission) dyeswere also purchased from Molecular Probes(Eugene, OR).

Skin samples were obtained boththrough a university tissue-sharing pro-gram and from an outside donor source.Rat skin was excised fresh immediatelyfollowing sacrifice at the LARR facilityon the Texas A&M campus, includingskin-only and skin-fat-muscle samples.Pig skin, including a thick (~3cm) layer offat, was obtained frozen from K&C MeatProcessing (Navasota, TX).

ProceduresSolution StudiesFITC-dextran and TRITC-Con A weredissolved in pH 7.2 buffer made

CCD

Monochromator

Collection Fibers

Excitation Fiber

Cell Holder

Cuvette

ChuckSMAConnector

Fiber Positioner

MicroscopeObjective

BP Filter

Lamp

1. Fiber-optic fluorescence spectrometer for in vivo use.

The results indicate

that a polynomial

describes the

nonlinear relationship

between fluorescence

intensity and

concentration

rather well.

38 IEEE ENGINEERING IN MEDICINE AND BIOLOGY November/December 2000

in-house. Glucose was also dissolved inbuffer from the same stock. Aqueoussensor solutions of 1 mL were mixed forTRITC-Con A to FITC-dextran mass ra-tios of 250:2.5, 250:1.5, 250:4, 750:7.5,and 62.5:0.625 µg/mL.

The custom-designed optical systemdescribed above was used for all fluores-cence measurements. In each experiment,the complete spectrometer system wasturned on and allowed to run for a mini-

mum of 30 min prior to use. The detectortemperature was maintained at −80 ºCthroughout the duration of data collection.Methacrylate cuvettes, placed within thecustom-built sample chamber, were usedto contain samples for all in vitro tests.Aqueous sensor solutions were stirredcontinuously with a 5 mm bar propelledby a magnetic stirrer to which the samplechamber was mounted. Integration timesfor the spectra varied between data sets,

though all exposures were less than 1 sec.Excitation light was blocked betweenscans to minimize photobleaching effects.

All samples were scanned under simi-lar conditions during incremental additionof glucose at a level of 10 mg/dL for eachstep (2.5 µL of 4000 mg/dL stock solu-tion). Samples were scanned immediatelyupon addition of glucose, with three con-secutive spectra recorded for each sample,except for 250:2.5, where only one scanwas taken per sample. Once it became evi-dent that the response to glucose had flat-tened, the increments were increased to 40mg/dL (10 µL) per step.

A final sensor solution of 2 mL for aTRITC-Con A:FITC-dextran mass ratioof 250:2.5 was generated. This mixturewas chosen because of the greater slope ofthe curve for intensity ratio versus con-centration. Glucose was added at a finerstep size, 2 mg/dL (2.5 mL of 1600mg/mL stock), allowing analysis of preci-sion versus sensitivity.

CalibrationSince the ultimate goal of this project is topredict glucose levels based upon fluores-cence signals, a calibration model isneeded to convert the detected light infor-mation into glucose concentration. Sev-eral different calibration procedures weretested toward this end. The data used forthese studies were collected on the250:2.5 sample with 2 mg/dL incrementsin glucose levels. For the purposes of thisstudy, the data were broken into two re-gions, based on the appearance of a linearrelationship: a lower concentration range(0-63.4 mg/dL) and the full range(0-182.5 mg/dL). Within these sets, thedata were further divided into calibration(75%) and validation (25%) sets, the for-mer for use in developing the calibrationmodels and the latter as an independenttest of the constructed models.

Univariate calibration was performedwith least-squares and polynomial regres-sions using single-point intensity and ra-tio values computed from the spectra.Another method for producing relativevalues of the two fluorescence emissionpeaks was also used: the experimentalspectra were regressed onto well-definedpure spectra of the dyes. Since the spec-trum (Y) of a mixture is composed of thefluorescent dye emission curves (f =FITC, t = TRITC) and reflected excitationlight profile (r), it can be written as a lin-ear combination of the three as: Y = αf +βt + γr. Since f, t, and r can be measured

(a)

(b)

2. Pictures showing (a) implantation sites and (b) monitoring procedure.

independently, the coefficients α, β,and γmay be obtained by least-squares regres-sion. These coefficients represent the con-tribution of each component to thespectrum and may therefore be used asmeasures of the relative intensity of FITCversus TRITC emission. Ratios of the co-efficients (α/β) produce a scaling com-pensation nearly identical to that achievedwith peak ratios. Furthermore, it is logicalthat the regression coefficients may bebetter suited to in vivo measurements be-cause they can directly compensate for re-flected light and skin autofluorescencespectra. Thus, while intensity ratios canbe adversely affected by weak sensor sig-nals superimposed upon stronger contri-butions from skin, the regressioncoefficients should be less sensitive tosuch conditions.

Both partial least-squares (PLS) [19]and parallel (PAR) [20] calibrations wereperformed with multipoint data, includingfull spectra and wavelength-selected re-gions. Wavelength selection was appliedto fluorescence spectra for two specificpurposes: 1) to determine an optimal pairof wavelengths to use for a ratio, giving thelowest prediction error in univariate cali-bration; and 2) to identify the best spectralregions for use with PLS and PAR calibra-tion. Details on the selection algorithms areavailable in the literature [21-23].

The results of the models were com-pared in terms of standard calibration(SEC) and prediction errors (SEP), meanpercent error (MPE), correlation coeffi-cient (R), and the offset (a) and slope (b)for the best-fit line between predicted andactual concentration values (a = 0, b = 1indicates a perfect fit) [19]. The percent-age of determinations falling within the15% “acceptable” margin were also tabu-lated for the validation data in each trial.The SEC was computed for the calibra-tion data only, while the rest of the mea-sures were calculated using only thevalidation data.

Excised Tissue ExperimentsSolutions of fluorescent dyes were in-jected into rat dermis to determine theability to obtain accurate spectral datathrough tissue. Spectra were collected byplacing the probe tip at the skin surfaceabove the injection site and moving theprobe while scanning to determine thepoint of strongest signal. Sites where sur-face leakage was observed were excludedfrom the study to avoid artificially high

fluorescence measurements due to liquidon top of the sample.

Microsphere injections were per-formed on the thicker pig skin, as thesetissue samples were larger and easier tomanipulate. A four-inch square slab of tis-sue, including skin and fat, was excisedfrom the back of a porcine subject and tenflaps of various depths cut from the skinsurface with a scalpel. A 5 mL aliquot of aprescanned sphere mixture was placedinto each shallow gouge beneath the skin

flap. The flaps were then pushed back intoplace and fluorescent spectra of the im-planted spheres collected with the probehead placed directly on top of the flap.

Animal StudiesAn animal-use protocol was approvedthrough the Texas A&M University Insti-tutional Review Board for the rat studies.Three rats were obtained from the TAMULARR facility, weighed, and anesthetizedaccordingly using ketamine and xylazene.


1.0

0.8

0.6

0.4

0.2

500 550 600Wavelength (nm)

Nor

mal

ized

Inte

nsity

Glucose

3. Normalized fluorescence emission spectra of 250:2.5 TRITC-Con A:FITC-dex-tran mixture for 0-400 mg/dL glucose.

1.25

1.00

0.75

0.50

0.25

0.00

0 100 200 300 400

Glucose Concentration (mg/dL)

Rat

io

Peak Intensity Ratio

Regression Coefficient Ratio

4. Sensor response curves showing fluorescence intensity and regression coefficientratios versus glucose concentration.

Injections were administered intramuscu-larly into the quadriceps muscle. Depth ofanesthesia was assessed by pinching frontand back paws prior to implantation. Theanimal’s eyes were rubbed with ointmentto prevent drying and the animals wereplaced on heat pads to maintain body tem-perature. Core temperatures were moni-tored periodically and temperatures werecontrolled between 93 and 95 ºF for theduration of the experiments.

Three 2 mL samples of the commer-cial polystyrene fluorescent micro-spheres were mixed in buffer containing20% glycerin to obtain three different ra-tios of emission peak intensities. Thesamples were scanned in vitro to confirm

appropriate spectral features before be-ing drawn through 30 G needles into 1mL insulin syringes. Samples were con-tained in the syringes until injection, atwhich time they were shaken to remixsettled beads. Intradermal injectionswere made dorsally immediately follow-ing anesthesia. Each of the three sampleswas injected three times per animal, re-sulting in a total of nine injections per ratand 27 overall. Each injection involvedpinching the skin, sliding the needlenearly parallel to the skin surface for der-mal placement, and injection of 0.05 mLof fluid. During injections for the first rat,it was observed that the spheres were notwell mixed; thus, the syringes wereshaken until the solution was visibly ho-mogeneous, and injections resumed. Im-plantation sites were circled with apermanent marker prior to injection, asshown in Fig. 2(a), to allow post-experimental collection of the tissues.

Rats were placed on a clean table nextto the fluorescence spectrometer andwithin reach of the fiber-optic probe. Theprobe head was placed in contact with theskin superficial to the implants, as shownin Fig. 2(b), and moved until a strong fluo-rescence signal was observed. Three spec-tra were collected for each site and theprobe was repositioned between scans.The animals were euthanized with CO2 af-ter the experiments were completed, andtissue samples were removed. Each of the27 sites was removed with surgical scis-sors, fixed in formalin, and delivered tothe Veterinary Histology department for

sectioning. The tissue samples were slicedinto 5 µm sections and mounted onto mi-croscope slides for examination.

Results

Solution StudiesNormalized spectra from the sensor so-

lutions collected in vitro are shown in Fig. 3.The fluorescence of the FITC emissionpeak increases with glucose concentrationas dextran molecules are displaced fromCon A by glucose and FRET diminishes.The data exhibit an approximately linear in-crease with glucose level for the concentra-tion range of 0-63.5 mg/dL, as seen fromFig. 4. After this point, the sensitivity is de-creased, resulting in a bending of the curvetoward zero slope. The potential for predic-tion of unknown glucose levels was evalu-ated with the 250:2.5 FITC-dextran:TRITC-Con A data collected at 2 mg/dL in-tervals. Univariate regressions, includingstraight-line and polynomial fitting, wereattempted with raw peak intensity values,peak intensity ratios, regression coeffi-cients, and transformed data. Of these, mod-els using intensity ratios and regressioncoefficients produced the lowest predictionerrors. Multivariate calibration was per-formed on the raw and intensity-scaledspectra using PLS and PAR.

The best results of each calibrationstudy, for each concentration range, aregiven in Tables 1 and 2. The “acceptable”heading refers to the percentage of predic-tions that fall within 15% of the true level,as stipulated by the ADA [24]. Slope, in-


Table 1. Calibration Statistics for the 0-63.5 mg/dL Range

Method SEC(mg/dL)

SEP(mg/dL)

MPE(%)

R Intercept(mg/dL)

Slope Acceptable(%)

Linear 0.81 0.89 3.8 0.999 0.38 0.98 95.8

PLS 0.59 0.83 3.62 0.999 0.56 0.98 95.8

PAR 0 2.18 10.5 0.999 1.01 0.99 87.5

Table 2. Calibration Statistics for the 0-182.5 mg/dL Range

Method SEC(mg/dL)

SEP(mg/dL)

MPE(%)

R Intercept(mg/dL)

Slope Acceptable(%)

Linear 14.9 14.7 49.0 0.960 4.24 0.94 55.2

Polynomial 3.2 3.0 4.1 0.998 -0.51 1.01 95.7

PLS 9.2 10.1 35.6 0.981 4.28 0.96 75.4

PAR 0 1.12 7.0 0.999 0.91 0.98 83.3

The best prediction

results were obtained

using polynomial

regression with

regression coefficient

ratios.

tercept, and R values are for the regres-sion of predicted values onto referencevalues of the validation set only; there-fore, a unity slope and zero offset areideal, and R should approach unity forproper calibration.

The results indicate that a polynomialdescribes the nonlinear relationship be-tween fluorescence intensity and concen-tration rather well. Both intensity andcoefficient ratios show a significant im-provement over the straight-line regres-sion, particularly for the 3rd- and5th-order polynomial functions, whichpredict that < 5% error and > 95% of vali-dation points fall within the acceptablerange. The best model for use with thecurrent sensor chemistry is the 3rd-orderpolynomial regression. None of the othercalibration methods come close in predic-tion accuracy for the full glucose range. Itappears from this information that, if thetotal sensor range can be increased tocover the full clinically important region(50-400 mg/dL), viable prediction modelscan still be obtained from nonlinear datawith relative ease.

As a final step in this phase of the re-search, data from the different ratio exper-iments were subjected to the polynomialregression to test the potential for deter-mining glucose at higher concentrations.A correlation plot was constructed forpredictions using the 62.5:0.625 ratio of

TRITC-Con A:FITC-dextran. This dataset was chosen because of the extent towhich the glucose response was observed,reaching almost to 600 mg/dL and the lesssevere vertex of the curve. A 5th-orderpolynomial was used, selected due to theleast percent error achieved over other or-ders. A correlation plot for this calibrationis shown in Fig. 5, where it is seen that thepredictions are nearly within the accept-able limits for this range. The statistics ac-companying this graph are SEC = 24.4mg/dL, SEP = 24.2 mg/dL, MPE = 22.5%,R = 0.989, intercept = 10.14 mg/dL, slope= 0.919, and acceptable = 74.1%. Whilethese numbers are not unreasonable, fur-ther improvements must be made beforean implanted sensor with a similar re-sponse profile can be expected to provideaccurate determinations.

Wavelength selection was also appliedto determine the best data points forstraight-line regression and the best spec-tral areas for multivariate calibration. Inthe first case, all possible wavelengthpairs were tested. For the multivariatemethods, the selection algorithms de-scribed above were applied to find a sub-set of data points achieving the highestprediction accuracy. The best pair ofwavelengths selected for linear regressionwas 525 and 600 nm, which are located onthe shoulders of the emission bands. Fur-ther examination revealed that all combi-

nations of data points found on the emis-sion bands produced nearly equivalentprediction errors. Thus, the choice ofpoints for the ratio calculation is not criti-cal in the univariate case. However, thesefindings do point out a potentially usefulbit of information, in that shoulder pointscan be used successfully for prediction. Ifthe detected signals are sufficientlystrong, this may allow the use of longerwavelength data, away from excitationpeaks, and so reduce the effects of scat-tered excitation light.

Fluorescence MicrosphereExperiments

Fluorescence spectra of the cus-tom-made sensor spheres were previouslyrecorded using a commercial fluorometer.The beads were immersed in glucose so-lutions of 0, 200, 400, 600, 800, and 1000mg/dL and scanned during excitation with488 nm light, indicating a similar re-sponse for the chemistry in an aqueousmilieu and encapsulated in the PEGspheres [15]. In fact, the response pos-sessed a greater linear region, up to 600mg/dL, compared to that seen in the solu-tions studied here. Despite these promis-ing results, the reverse response has notyet been realized in reasonable time peri-ods [15]. Because of these obstacles, itwas impractical to obtain enough data tothoroughly test the sensitivity and accu-racy of the beads. Therefore, the opticswere tested using excised tissue and invivo using solutions and static sphereswith known fluorescence properties.


600

500

400

300

200

100

00 200 400 600

Actual Concentration (mg/dL)

Pre

dict

ed C

once

ntra

tion

(mg/

dL)

Calibration Data

Validation Data

5. Glucose prediction using peak ratios and 5th-order polynomial.

While the actual

sensor implants are

still under

development, in vivo

experiments have been

conducted using

“simulated” sensors.

Due to the current limitations of thehydrogel spheres, skin studies were per-formed using both fluorescent solutionsand spheres without glucose sensitivity.Spectra of FITC- and TRITC-dextranwere collected prior to mixing and injec-tion, as were spectra of plain skin and thesample holder. Normalized in vitro spec-tra of FITC- and TRITC-dextran areshown in Fig. 6, along with skin spectraobtained by placing the probe on a sectionof rat skin, and a spectrum obtained fromthe plate used to mount the skin. The simi-larity between the skin and plate samples

is obvious, caused by the major influenceof the excitation light. However, the skinalso exhibits weak fluorescence, as evi-denced by the increased signal at longerwavelengths. A small bump in the spectraat ~546 nm is attributed to the mercuryline of the excitation source.

The four components shown in Fig. 6almost completely account for the finalspectra acquired from injected dyes. Theagreement between experimentally ac-quired spectra of injected dye and spectraestimated from the four contributing com-ponents is excellent (< 2% error), suggest-

ing that a regression of spectra onto purecomponents may allow determination ofrelative peak intensities independent ofthe excitation light scattering that willlikely vary from site to site. Because thesensor response for the encapsulated as-say has not yet been perfected, experi-ments to test this hypothesis continuedwith “simulated” sensors. Two sets of 15µm diameter polystyrene microsphereshaving fluorescence peaks positionednear those of FITC and TRITC were usedfor these studies.

Figure 7 includes spectra of thesespheres as collected in vitro. The yel-low-green dye does not have a well-de-fined peak, but it reaches a maximumaround 515 nm. The orange dye, on theother hand, possesses a distinct peak at~560 nm. The spheres were mixed to-gether at varying mass ratios to obtain dif-ferent peak ratios, simulating the changesseen in the FRET-based assay with differ-ent glucose levels.

In order to prove that transdermal mea-surements can be correlated with in vitromeasurements, three mixtures containingdifferent peak ratios were generated.These mixtures were used to simulate 0,500, and 1000 mg/dL concentrations,where the larger ratio values correspondto higher concentrations. In vitro scans ofthese solutions are shown in Fig. 8, whereit is seen that the 515 nm:560 nm peak ra-tios move from 0.33 to 0.88 to 1.43 overthe three different mixtures. These are in anearly perfect straight line, simulating alinear glucose response curve. The chal-lenge, then, is to be able to make similarmeasurements in vivo.

Fluorescence regression coefficientratios for in vivo spectra are plotted inFig. 9. The ratio values are highly vari-able—not only do they show poor repeat-ability, they also deviate from the in vitroratios. There does appear to be a weakcorrelation between the in vitro and aver-age in vivo peak ratios for two sets ofdata, but the in vivo measurements have asmaller slope. This was already noted tobe a problem likely caused by scatteringnear the yellow-green emission peak andthe excitation of different coloredspheres. The slope difference can becompensated with calibration, but thespread in the measurement points is agreater concern. The overlap is so severethat one set of samples (i.e., one “concen-tration”) cannot be statistically separatedfrom the others. The groupings of the ra-tio values by site, distinguished in the


1

0.8

0.6

0.4

0.2

0500 520 540 560 580 600 620 640

Wavelength (nm)

Nor

mal

ized

Inte

nsity

Plate

Skin

FITC-Dextran TRITC-Dextran

6. Spectra of dyes, skin, and mounting plate.

250000

200000

150000

100000

50000

0500 550 600

Wavelength (nm)

Inte

nsity

Yellow-Green

Orange

7. In vitro fluorescence spectra of 15 µm diameter beads.

plot by different symbols, do not indicateany special dependency on the implanta-tion site, such as increased relative scat-tering or absorption. It is notable that thevariation in ratios appears to increasewith the simulated concentration, whichcorresponds to a relative increase in fluo-rescence at the yellow-green emission.The increasing variation is likely due tothe greater scattering at lower wave-lengths, which then makes the measure-ments more dependent upon depth.

A general dependence of spread inratio measurements on implantationdepth was noted. For example, given asolution that was injected in threeplaces, the overall measured ratio wasgenerally less precise when a wide rangeof depths were encountered. In addition,the ratios tend to have smaller values fordeeper injections. While this result issomewhat intuitive, it underscores theprevious statement of the need for cali-bration after implantation because ra-tios are clearly different in vivo versusin vitro.

Another explanation for the large vari-ance in the in vivo ratio measurements isthe inhomogeneity of the injected sample.Because of the brightness of themicrospheres, the suspensions were di-lute. The small volume injected at eachsite was not composed of equal numbersof the two colors of spheres. Thus, a broaddistribution around the mean was ob-served. These experiments will be re-peated in the future using custom-mademicrospheres with encapsulated dye mix-tures of different color ratios.

It should be noted that the spheres usedin these studies are approximately 200times brighter than the (custom-made)PEG spheres, which contained the samedye concentration used in the sensorspheres. Thus, the fluorescence spectracollected from implanted sensor sphereswill not be nearly as strong as those ob-tained in this study if the current dye con-centrations are retained.

The final step in analyzing the in vivodata was to back-calculate the requiredconcentration changes in the sensorspheres to allow in vivo monitoring. Thiswas accomplished by taking the highestand lowest regression coefficients calcu-lated for the yellow-green emission peak,and reconstructing the expected spec-trum from FITC. Because the commer-cial polystyrene spheres are 200 timesbrighter than the FITC, the coefficientswere divided by 200 then multiplied by

the FITC spectrum in the reconstruction.Figure 10 contains a plot showing the es-timated spectra for the best and worstcases. In the worst case, the expectedspectrum looks exactly like that of skin.In the best case, a TRITC emission is ob-served, but the FITC emission is ob-scured by the skin spectrum. Because theskin spectrum represents noise at theFITC peak, it is desirable to have theFITC emission at least ten times strongerthan the skin spectrum. In order toachieve this level of FITC fluorescence,the concentration of the fluorophore inthe sensor spheres must be increased byat least a factor of 26. While this may notbe feasible due to the self-quenching ef-fects and solubility limitations at highconcentrations, this requirement may po-

tentially be relaxed by using other detec-tion methods, longer wavelength dyes, ortwo-photon excitation.

ConclusionA fiber-optic probe was fabricated for

delivery and collection of light to an im-planted sensor. The probe and associatedoptical system has high sensitivity and al-lows for flexible, remote measurements.The system was used to collect data withsensor solutions similar to what is envi-sioned for the implantable system, and therecorded spectra were processed in an at-tempt to predict glucose concentrationsfrom fluorescence measurements. De-spite the availability of sophisticatedmultivariate calibration methods, the bestprediction results were obtained using


120000

80000

40000

0

500 550 600

Wavelength (nm)

Inte

nsity

0 mg/dL

500 mg/dL

1000 mg/dL

8. Fluorescence spectra of bead mixtures.

4

3

2

1

0

−10 200 400 600 800 1000

InV

ivo

Coe

ffici

ent R

atio

Simulated Concentration (mg/dL)

9. Regression coefficient ratios for in vivo spectra.

polynomial regression with regression co-efficient ratios. The measurements are farfrom perfect, but they do meet the require-ments for clinical use and show promisefor further improvement in the future.

In vivo experiments were conductedusing “simulated” sensors that had con-stant fluorescence characteristics. Resultsof implantation in excised skin and in liveanimals indicate that, while very strongfluorescence signals were measured forimplanted particles, the current sensorchemistry must be improved for adequateresolution. In addition, difficulty inachieving repeatable injection depths andhomogeneity of mixtures containing dif-ferent colored spheres confounded at-tempts to make true quanti tat ivemeasurements.

AcknowledgmentsThe authors thank Dr. Cliff

Spiegelman and Dr. Lihong Wang fortheir gracious help in this project. Thiswork was supported by fellowships fromthe Whitaker Foundation (MJM) and Tex-aco Corp. (RJR) as well as grants from theWhitaker Foundation and the NationalScience Foundation (# BES-9908439).

Michael J. McShane re-ceived the B.S. degreewith Magna Cum Laudehonors in 1994 and thePh.D. in 1999, both inbiomedical engineeringfrom Texas A&M Uni-versity. He started as anassis tant professor

within the Biomedical Engineering Pro-

gram at Louisiana Tech University inSeptember of 1999. He is also on the fac-ulty of the university’s Institute forMicromanufacturing (IfM).

Since joining Louisiana Tech Univer-sity, Dr. McShane is the codirector of theBioMEMS Laboratory at the IfM. His pri-mary research interests focus on the use ofinstrumentation for physical and chemicalbiomedical measurements, particularlyusing optical sensors. His other researchinterests include tissue engineering andimaging. He is a member of the IEEE En-gineering in Medicine & Biology Society,BMES, the SPIE Biomedical OpticsGroup, the Society for Applied Spectros-copy, and the Tau Beta Pi EngineeringHonor Society.

Ryan Russell received aB.S. degree in AerospaceEngineering from TexasA&M University. Aftergraduating in 1994, heworked in the missionoperations directorate atNASA-Johnson. He wasbeen working toward a

Ph.D. in chemical engineering at TexasA&M University since 1997. His researchinterests include flourescent biosensing,polymer thin films, and computer simula-tion. He is a student member of the IEEEEMBS.

Michael Pishko received his B.S. and M.S.degrees in chemical engineering from theUniversity of Missouri-Columbia and hisPh.D. in 1992 from the University of Texasat Austin in the area of implantable

biosensors. He was alsoa postdoctoral associateat MIT and instrumentalin the startup of twocompanies, both in thearea of glucose sensorsfor diabetes manage-ment. He joined thechemical engineering

faculty at Texas A&M in January of 1997as an assistant professor. He is currentlyengaged in interdisciplinary research inthe biomaterials and biosensors area. Hisresearch is currently funded by theWhitaker Foundation, the Juvenile Diabe-tes Foundation International, NSF,NASA, and the US Army. He has receivedan NSF Career award, the OutstandingYoung Scientist award from the HoustonSociety for Engineering in Medicine andBiology, and was named an Alfred P.Sloan Research Fellow.

Gerard L. Coté re-ceived the B.S. degreein electrical engineeringwith honors in 1986from Rochester Insti-tute of Technology, theM.S. degree in 1987,and the Ph.D. in 1990,both in bioengineering

from the University of Connecticut. Fol-lowing a five month postdoctoral study atthe University of Connecticut, he startedas an assistant professor within the Bio-medical Engineering Program at TexasA&M University in May of 1991. He waspromoted to associate professor, with ten-ure, in September 1997. He is also a clini-cal assistant professor at the University ofTexas Medical Branch in the Departmentof Ophthalmology & Visual Sciences.

Since joining Texas A&M, Dr. Cotéhas created and is the Director of the Opti-cal Biosensing Laboratory within the Bio-medical Engineering Program. Hisprimary research interests focus on theuse of optics and fiber optics for medicaldiagnostics and biosensing. He has an in-terest in the development as well as thetheoretical modeling and analysis of opti-cal sensors for both physical and chemicalbiomedical measurements. His other re-search interests include bio-optical andbiomedical imaging. He is a member ofIEEE Engineering in Medicine & BiologySociety, the IEEE Lasers & Electro-Op-tics Society, the SPIE Biomedical OpticsGroup, the Biomedical Engineering Soci-ety, the American Society for EngineeringEducation, and the Tau Beta Pi Engi-neering Honor Society.


1.0

0.5

0.0

Inte

nsity

500 550 600Wavelength (nm)

Worst Case

Best Case

Best Case*26

10. Estimated emission spectra from implanted sensors.


Address for Correspondence: MichaelJ. McShane, Biomedical EngineeringProgram, Louisiana Tech University,P.O. Box 10348, 600 W. Arizona St.,Ruston, LA 71272. Tel: +1 318 2573191. Fax: +1 318 257 2562. E-mail:[email protected].

References1. Guyton AC: Textbook of Medical Physiology,7th ed. Philadelphia, PA: W.B. Saunders, 1986.2. National Diabetes Clearinghouse: Diabetesstatistics. NIH Publication No. 99-3926, March1999.3. Diabetes Research Group: The effects of in-tensive treatment of diabetes on the developmentand progression of long-term complications in in-sulin-dependent diabetes mellitus. N Engl J Med329: 977-986, 1993.4. Lakowicz JR and Maliwal B: Optical sensingof glucose using phase-modulation fluorimetry.Analytica Chimica Acta 271: 155-164, 1993.5. Ballerstadt R and Schultz JS: Competi-tive-binding assay method based on fluorescencequenching of ligands held in close proximity by amultivalent receptor. Analytica Chimica Acta345: 203-212, 1997.6. Tolosa L, Szmacinski H, Rao G, andLakowicz JR: Lifetime-based sensing of glucoseusing energy transfer with a long lifetime donor,”Analytical Biochem 250: 102-108, 1997.7. Tolosa L, Malak H, Raob G, and LakowiczJR: Optical assay for glucose based on the lumi-nescence decay time of the long wavelength dyeCy5TM. Sens Actuators B, Chem 45: 93-99,1997.

8. Meadows DL and Schultz JS: Design, manu-facture, and characterization of an optical-fiberglucose affinity sensor based on an homogeneousfluorescence energy-transfer assay system.Analytica Chimica Acta 280: 21-30, 1993.

9. Meadows D and Schultz JS: Fiber-opticbiosensors based on fluorescence energy-transfer.Talanta 35: 145-150, 1988.

10. Lakowicz JR: Principles of FluorescenceSpectroscopy. New York: Plenum, 1983.

11. Clark LC: Continuous measurement of circu-lating glucose using the transcutaneous pO2 elec-trode. Birth Defects 15: 39-42, 1979.

12. Quinn CP, Pishko MV, Schmidtke DW,Ishikawa M, Wagner JG, et al.: Kinetics of glu-cose delivery to subcutaneous tissue in rats mea-sured with 0.3-mm amperometric microsensors.Amer J Physiol-Endocrinol Metabolism 32:E155-E161, 1995.

13. West J and Hubbell J: Photopolymerizedhydrogel materials for drug-delivery applications.Reactive Polymers 25: 139-147, 1995.

14. Russell RJ, Gefrides CC, McShane MJ,Coté GL, and Pishko MV: A fluorescence-basedglucose biosensor using concanavalin A and dex-tran encapsulated in a poly(ethylene glycol)hydrogel. Analytical Chem 71: 3126-3132, 1999.

15. Russell RJ, Pishko MV, Gefrides CC, andCoté GL: A fluorescent glucose assay usingpoly-L-lysine and calcium alginate micro-encapsulated TRITC-succinyl- Concanavalin Aand FITC-dextran. Presented at the 20th AnnualInternational Conference of the IEEE-EMBS,Hong Kong, 1998.

16. McShane M, Rastegar S, Pishko M, andCoté G: Monte Carlo modeling for implantable

analyte sensors. IEEE Trans Biomed Eng, to bepublished.

17. McShane MJ, Russell RJ, Pishko MV, andCoté GL: Optical system for implantableanalyte sensors. Presented at 21st Annual Inter-national Conference of the IEEE-EMBS, At-lanta, GA, 1999.

18. McShane MJ, O’Neal DP, Russell RJ,Pishko MV, and Coté GL: Progress towardimplantable fluorescence-based sensors for moni-toring glucose levels in interstitial fluid. Presntedat SPIE International Biomedical Optics Conf,San Jose, CA, 2000.

19. Martens H and Næs T: Multivariate Cali-bration. New York: Wiley & Sons, 1989.

20. Spiegelman CH, McShane MJ, and CotéGL: A transparent tool for seemingly difficultcalibrations: The parallel calibration method. An-alytical Chem 72:135-140, 2000.

21. McShane MJ, Coté GL, Spiegelman CH:Variable selection in multivariate calibration of aspectroscopic glucose sensor. Applied Spectros-copy 51: 1559-1564, 1997.

22. Spiegelman CH, McShane MJ, Goetz MJ,Motamedi M, Yue QL, and Coté GL: Theoreti-cal justification of wavelength selection in PLScalibration: Development of a new algorithm. An-alytical Chem 70: 35-44, 1998.

23. McShane MJ, Cameron BD, Coté GL,Motamedi M, and Spiegelman CH: A novelpeak-hopping stepwise feature selection methodwith application to Raman spectroscopy.Analytica Chimica Acta 388: 251-264, 1999.

24. Anonymous: Consensus statement onself-monitoring of blood-glucose. Diabetes Care10: 95-99, 1987.

Glucose monitoring using implanted fluorescent microspheres

Documents