Nocturnal Hypoglycemic Alarm Based on Near-Infrared Spectroscopy

Jan 3, 2019 - Given an alarm threshold (e.g., 3.0 mM), a database of calibration differential ... In this work, the alarm algorithm was tested in two ...
6 downloads 0 Views 1MB Size
Download PDF
Article pubs.acs.org/ac

Cite This: Anal. Chem. XXXX, XXX, XXX−XXX

Nocturnal Hypoglycemic Alarm Based on Near-Infrared Spectroscopy: In Vivo Studies with a Rat Animal Model Sanjeewa R. Karunathilaka,† Mark A. Arnold, and Gary W. Small* Department of Chemistry & Optical Science and Technology Center, University of Iowa, Iowa City, Iowa 52242, United States

Downloaded via UNITED ARAB EMIRATES UNIV on January 12, 2019 at 01:36:04 (UTC). See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

S Supporting Information *

ABSTRACT: A noninvasive method for detecting episodes of nocturnal hypoglycemia is demonstrated with in vivo measurements made with a rat animal model. Employing spectra collected from the near-infrared combination region of 4000− 5000 cm−1, piecewise linear discriminant analysis (PLDA) is used to classify spectra into alarm and nonalarm data classes on the basis of whether or not they correspond to glucose concentrations below a user-defined hypoglycemic threshold. A reference spectrum and corresponding glucose concentration are acquired at the start of the monitoring period, and spectra are then collected continuously and converted to absorbance units relative to the initial reference spectrum. The resulting differential spectra correspond to differential glucose concentrations that reflect the differences in concentration between each spectrum and the reference. Given an alarm threshold (e.g., 3.0 mM), a database of calibration differential spectra can be partitioned into two groups containing spectra above and below the threshold. A classification model is then computed with PLDA. The resulting model can be applied to the differential spectra collected during the monitoring period in order to identify spectra whose corresponding glucose concentrations lie in the hypoglycemic range. In this work, the alarm algorithm was tested in two single-day studies performed with anesthetized rats. Glucose concentrations spanned the range of 1.6 to 13.5 mM (29 to 244 mg/dL). For both rats, the alarm algorithm performed well. On average, 87.5% of alarm events were correctly detected, and the occurrence of false alarms was 7.2%. False alarms were restricted to times when the glucose concentrations were very close to the alarm threshold rather than at random times, thus demonstrating the potential of the approach for practical use.

S

we are seeking applications that have more relaxed requirements regarding the portability of the instrumentation. One such application, and the focus of the current paper, is the development of a nocturnal hypoglycemic alarm. Nocturnal hypoglycemia is the occurrence of low blood glucose levels during sleep and is an important problem among diabetics who use insulin as part of their treatment regimen. We envision that a noninvasive hypoglycemic alarm would continuously monitor a patient’s glucose level during the sleep period, sounding an alarm if the glucose level decreases below a set threshold. Current invasive continuous glucose monitors18 (CGMs) that are implanted in the subcutaneous tissue have similar alarm functions that can be used to signal hypoglycemia.19 Research has also been performed to use the output of a CGM specifically to predict nocturnal hypoglycemia in advance of its occurrence.20,21 We have recently completed three studies that constitute important steps toward the development of a noninvasive hypoglycemic alarm. The first reported on the development of

everal technologies are currently being evaluated for measuring blood glucose noninvasively (i.e., without the physical collection of a blood sample). The approaches include the use of vibrational transmission or reflectance spectroscopy to probe dermis tissue,1,2 Raman scattering measurements,3,4 interrogations of tear fluid with glucose-sensitive reagents,5 and the determination of glucose in interstitial fluid extracted by iontophoresis.6 Our laboratories have a long-standing interest in the development of near-infrared (near-IR) transmission measurements for the determination of blood glucose.7 The approach used in our work is to pass near-IR light through dermis tissue and to extract quantitative glucose information from the resulting measured spectra. Our analysis focuses on the combination region of the near-IR between 4000 and 5000 cm−1. This spectral region contains two distinct glucose C−H stretch−bend combination bands at 4300 and 4400 cm−1 that lie in the trough between two broad and intense absorption bands of water. There are many challenges to the development of a noninvasive blood glucose sensor, and we have taken an incremental approach that addresses the various issues related to the implementation of a successful measurement.8−17 Before a truly miniaturized, portable glucose sensor can be realized, © XXXX American Chemical Society

Received: July 31, 2018 Accepted: January 3, 2019 Published: January 3, 2019 A

DOI: 10.1021/acs.analchem.8b03437 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

Figure 1. (A) Comparison of a spectrum of rat tissue with a similar measurement made on the back of the human hand. The spectral shapes are similar. Higher absorbance is noted in the rat spectrum because of increased light scattering. (B) Plot of 462 absorbance spectra for rat A skin tissue relative to air. Fat absorbance features can be seen at 4250 and 4350 cm−1, while the peak at 4600 cm−1 arises from skin tissue proteins.

while glucose excursions were induced through the infusion of glucose or insulin through catheters that were installed via a surgical procedure. Each rat was studied during a single-day experiment. A detailed description of the experimental protocol and photographs of the experimental setup are provided in the Supporting Information. The targeted measurement site on the rat was used to simulate noninvasive human measurements. The rat skin absorbance spectrum shown in Figure 1A reveals water as the main spectral contributor, with the edges of the spectrum increasing in absorbance in the direction of the intense water bands located near 3300 and 5200 cm−1.26 The peaks at 4250 and at 4350 cm−1 arise from fat absorbance in the skin tissue, and the peak at 4600 cm−1 corresponds to proteins, mainly keratin and collagen, that comprise the epidermis and dermis layers of skin. An absorbance spectrum for human skin on the back of the hand23 is also displayed in Figure 1A and shows a similar shape to the spectrum collected from the rat skin tissue. The similarities in the shapes of the spectra indicate that the main chemical components are present in reasonably similar amounts, with fat appearing to be a larger contributor to the rat spectrum. Olesberg et al. reported that the differences in

a general-purpose data analysis algorithm for threshold monitoring applications.22 In this study, the algorithm was tested with near-IR spectra of solutions of glucose in water along with a minimum of interfering background components. We demonstrated that we could classify spectra into groups corresponding to being above or below a concentration threshold. The second study described an effort to create in vitro samples that have a similar near-infrared spectral response to human tissue.23 In a follow-on study, we simulated the operation of the nocturnal alarm by use of this constructed tissue phantom.24 The work described in the current paper extends the testing of the nocturnal alarm methodology to the use of a rat animal model to implement true in vivo noninvasive blood glucose measurements.25 Glucose excursions were induced that mimicked the pattern of glucose changes that could occur during sleep. Glucose concentrations ranged from 1.6 to 13.5 mM (29 to 244 mg/dL), bracketing the 3.0 to 4.0 mM (54 to 72 mg/dL) range that corresponds to hypoglycemia.

■

EXPERIMENTAL SECTION Sequential near-infrared spectra were collected from the thin skin on the upper shoulder area of two Sprague−Dawley rats, B

DOI: 10.1021/acs.analchem.8b03437 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

to implement the alarm. Here, the first portion of the data acquired from the rat was used to train the classification model, and the latter portion of the data was used to test the methodology. Analysis of Rat A. Rat A was a male weighing ∼438 g. The surgery was performed, and the glucose transients for the hypoglycemic studies were performed on the same day. The data collection lasted 8.3 h and produced 462 spectra. Measured glucose concentrations ranged from 1.6 to 11.3 mM (29 to 204 mg/dL). The short-term noise for rat A was evaluated by computing 100% lines for pairs of consecutive single-beam spectra. These 100% lines were converted to absorbance, and the region of 4500−4300 cm−1 was fitted to a second-order polynomial baseline model to remove systematic components. This spectral region brackets the principal glucose absorption band at 4400 cm−1. The RMS noise was then computed about the polynomial fit to obtain the intrinsic measurement noise. The average RMS noise calculated for consecutive pairs of spectra collected at 128 scans was 31.5 μAU. The corresponding median value was 20.7 μAU. A plot of RMS noise as a function of spectral sequence number is included as Figure S3. The RMS noise values reported above are based on the ratio of spectra collected approximately 0.5 s apart. Thus, these values do not incorporate longer-term sources of spectral variation such as instrumental drift, temperature variation, or the effects of the pressure applied to the tissue by the skin interface. These additional factors clearly influence the overall signal-to-noise ratio of the spectral data set associated with each rat. To gain insight into changes with time in the gross chemical composition of the tissue within the skin interface, multiple linear regression was used to fit the 462 absorbance spectra shown in Figure 1B to pure-component spectra of 1 mm thick samples of water, collagen, keratin, and fat. Collagen and keratin are the two most prevalent proteins in epidermis and dermis tissue. The model also included intercept and linear slope terms to help fit the spectral baseline. This diagnostic procedure has been described previously.23 Over the 4800− 4200 cm−1 range, values of the coefficient of multiple determination (R2) for the fits all exceeded 0.9997. In the fitted models, the water and collagen terms had the largest regression coefficients, with the average ± standard deviation equaling 0.738 ± 0.005 mm for water and 0.596 ± 0.003 mm for collagen. Both coefficients changed rapidly when the tissue was first placed in the interface, with the water thickness decreasing and collagen thickness increasing as the tissue responded to being clamped in the interface. This can be interpreted as water being pushed out of the region between the sapphire rods of the interface and collagen filling in the vacated space. The total change in thickness was approximately 4% for both constituents. Figure S4 plots the % changes in the water and collagen coefficients relative to the first spectrum. The quality of the data was further evaluated by use of principal component analysis (PCA).31 The skin tissue absorbance spectra computed relative to an open-beam reference (Figure 1B) were mean-centered, and the PCA calculation was applied over the 4900−4200 cm−1 range. Hotelling’s T2 outlier detection metric32 was used to identify unusual observations. On the basis of these calculations, the first 16 spectra collected at the beginning of the experiment were removed from further analysis. Figure S5 plots the scores for the first two principal components.

the absolute absorbance values between human and rat tissue spectra are related to a higher degree of light scattering by the rat skin tissue.27 This can be seen by the higher absorbance values across the spectral range. The basic protocol for the data collection was designed to characterize the changes in skin spectra associated with hypoglycemia. Spectra were collected continuously for approximately 8 h. Figure 1B displays the spectra collected from rat A, presented as absorbance spectra computed relative to a mean open-beam background. Variation across the spectra is primarily a consequence of the tissue adjusting to the pressure applied by the skin interface.

■

RESULTS AND DISCUSSION Overview of Hypoglycemic Alarm Algorithm. The algorithm used to detect instances of hypoglycemia from continuously sampled near-IR spectra has been described previously22,24 and will be summarized here. A conventional invasive blood glucose measurement is made at the start of the monitoring period, and a reference tissue spectrum is collected concurrently. Subsequent sample spectra are computed in absorbance relative to this reference and are termed differential spectra, because the glucose concentration reflected in these spectra corresponds to the concentration difference relative to the reference value collected at the start of the monitoring period. In addition, any common background components between the sample and reference spectra will be suppressed to some extent by computing the spectral ratio, thereby reducing the contribution of the sample matrix. The alarm algorithm is implemented as a numerical pattern recognition problem, in which spectra are classified according to whether or not they represent differential glucose concentrations below a critical concentration threshold used to define hypoglycemia. The critical concentration (Ccrit) is defined as Ccrit = Calarm − Cref

(1)

where Calarm is a user-defined hypoglycemic threshold (e.g., 3.0 mM), and Cref is the reference glucose concentration measured at the start of the monitoring period. Assuming the subject is not hypoglycemic at the start of monitoring, Ccrit will be negative, and the classification problem is to identify spectra whose differential concentrations are more negative than Ccrit. Piecewise linear discriminant analysis (PLDA)28,29 is employed to perform the classification decision, and partial least-squares (PLS)30 is used to reduce the dimensionality of the input differential spectra before submission to PLDA. Each reduced dimensionality spectrum is termed a “pattern” that is indicative of the strength of the glucose signature. When supplied with a representative “training set” of patterns, the output of PLDA is a classification model. When a pattern of unknown classification is subsequently input to the model, the output is a discriminant score that indicates the class assignment associated with the pattern. Discriminant scores greater than zero correspond to assignment into the alarm data class, while negative discriminant scores denote the nonalarm class. In the ultimate implementation of this method with a human subject, the training set of differential spectra used to generate the classification model would be obtained by performing a glucose tolerance test, in which a series of spectra and reference glucose concentrations would be acquired. The resulting database of spectra would then be used going forward C

DOI: 10.1021/acs.analchem.8b03437 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry To assign an initial glucose concentration to each spectrum, the spectral collection time was linearly interpolated with the reference glucose concentrations and corresponding measurement times. Accurate concentration assignments, however, require the consideration of the time delay (lag) of glucose between the arterial blood used for the reference measurements and the interstitial fluid that dominates our near-IR tissue measurements.33 Reports of many research groups indicate that changes in the glucose level in arterial blood require 2−15 min before the corresponding concentration of glucose is observed in the interstitial fluid.34−39 The wide range of lag times reported in the literature illustrates the experimental challenges in making accurate determinations of this parameter and that a variety of physiological and metabolic factors can contribute to the rate of transport of glucose into the interstitial fluid. From a survey of the literature, there is no general agreement regarding the principal factors that affect lag time and under what conditions the lag time varies or remains constant. Because of this lack of guidance in the literature, we have adopted an empirical treatment of lag time in the current work. Using the interpolated concentrations and a range of lag times from 0 to 20 min in 1 min increments, a glucose concentration was assigned to each of the noninvasive spectra collected. A cross-validation PLS calculation was performed with the absorbance spectra, leaving out 10% of the data each cycle as the internal prediction set. The wavelength range was 4900−4200 cm−1, and 1−16 PLS latent variables (LVs) were used. The glucose concentration profile that gave the best standard error of cross-validation (SECV) value was assumed to correspond to the optimal delay time assignment. For each lag time, the SECV values were plotted with respect to the number of latent variables in the PLS model to determine the minimum SECV value. This plot is shown in Figure 2. The

Figure 3. Lag-corrected glucose concentration profile for rats A (A) and B (B). Each concentration profile was partitioned into a training set (red), a monitoring set (green), and a prediction set (blue). A horizontal line in each plot identifies the alarm threshold of 3.0 mM used for the prediction set. The horizontal line at 4.40 mM glucose in Figure 3B denotes the alarm threshold used with the monitoring set.

training data were used in building the classification model, while the monitoring data served as an initial set of external prediction data for use in evaluating and optimizing the model. Once the optimal model parameters were determined, the training and monitoring sets were combined to form a data set for use in computing the final classification model. The prediction set was withheld completely from the model development work and was reserved for use in testing the final classification model. Ratios of all the combinations of single-beam spectra in the training set were computed, and the resulting differential spectra and corresponding differential concentrations were used to optimize the wavenumber and number of PLS LVs that could be applied to the alarm classification of the monitoring set. The calculation of differential spectra for the training set produced 13 466 total differential spectra. A grid search was performed to identify the best wavenumber and latent variable combinations. The grid search was based on sliding a window of fixed spectral width in 25 cm−1 increments across the 4900−4100 cm−1 range. At each step,

Figure 2. Plot of SECV vs the number of PLS latent variables for rat A. The best lag time was 11 min based on the minimum SECV value of 0.63 mM.

delay time corresponding to the minimum SECV was taken as the optimal time required for the arterial blood to be equilibrated in the skin tissue matrix. For rat A, the optimal lag time was found to be 11 min, and the final glucose concentrations were assigned to each spectrum on the basis of this time. The glucose concentration profile was then partitioned into a training set (164 spectra), a monitoring set (83 spectra), and a prediction set (199 spectra) as shown in Figure 3A. The D

DOI: 10.1021/acs.analchem.8b03437 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

Figure 4. (A) Prediction glucose concentration profile for rat A. There were 27 alarm patterns and 172 nonalarm patterns. The alarm threshold of 3.0 mM is shown by the horizontal line. (B) Discriminant score plot for the prediction set with three replicate classifiers. The dashed line indicates a discriminant score of 0.0 that corresponds to the boundary between the alarm and nonalarm data classes. Red, blue, and green symbols denote the discriminant scores for the three replicate classification models. Dots, squares, and circles correspond to correct classifications, missed alarms, and false alarms, respectively.

scores computed with the optimal wavenumber−latent variable combination were partitioned into 8346 alarm and 22 035 nonalarm patterns. Three replicate classifiers were computed with the calibration PLS patterns. The prediction set contained 27 alarm and 172 nonalarm patterns. Figure 4A shows the prediction glucose concentration profile for rat A. The prediction result for the three replicate discriminants yielded 4, 2, and 0 missed alarms and 3, 8, and 14 false alarms. A committee decision rule was applied to these results (i.e., 2 out of 3 discriminant scores had to be positive for an alarm detection). When the committee rule was applied, there were 2 missed alarms out of 27 alarm patterns and 8 false alarms out of 172 nonalarms patterns. This corresponded to the successful detection of 92.6% of the alarm events with a false alarm rate of 4.7%. The discriminant score plot for the prediction set is shown in Figure 4B. The concentration trend in the prediction profile is clearly explained by the discriminant score plot. This dependence of discriminant scores on the glucose concentration further confirms the data classification by PLDA. A comparison between the discriminant score plot in Figure 4B and the prediction concentration profile in Figure 4A confirms that the false and missed patterns (circles and squares, respectively, in Figure 4B) correspond to concentrations close to the alarm threshold. Small differences in concentrations at the alarm threshold are challenging to detect, because the classification of these patterns is extremely sensitive to the positioning of the separating surface between the data classes. This factor explains why benefit is gained from training replicate classifiers and using them collectively to make the classification. Another factor that must be considered is imprecision in the assignment of the reference concentrations. As described previously, concentrations were assigned to each spectrum on the basis of interpolating the reference concentration measurements. Imprecision in these assignments can also lead to what appear to be missed or false alarms at the alarm threshold. Imprecision in the determination of the concentration lag can have a similar effect. Analysis of Rat B. Rat B was a male weighing ∼340 g. A nonsurvival surgery for the rat was performed, and the glucose

PLS models were constructed using 3−16 LVs. The starting width of 100 cm−1 was incremented in 25 cm−1 increments up to a maximum width of 700 cm−1. At each step, the crossvalidation procedure was applied, withholding 10% of the data for prediction. The top four wavenumber ranges were 4875−4175, 4850− 4175, 4900−4200, and 4900−4300 cm−1. For each of these ranges, plots of SECV vs the number of latent variables revealed a declining trace approaching 0.4 mM with little improvement past 14 LV. Accordingly, these four ranges were used together with 9−14 LVs to build classification models with PLDA. The PLS calculation was applied to all the wavenumber and PLS factor combinations to produce the corresponding sets of scores. The critical concentration of the monitoring set was then used to partition the differential spectra in the training set into alarm and nonalarm classes for use with PLDA. The reference concentration (Cref in eq 1), the first concentration in the monitoring set, was 7.12 mM, and the critical concentration (Ccrit) was −4.12 mM for an alarm threshold (Calarm) of 3.0 mM. There were 4234 alarm patterns and 9132 nonalarm patterns in the training set. Three replicate classifiers were computed with PLDA by changing the seeds of the random number generator used during the training process. Each classifier was based on three linear discriminant functions. The classifiers were applied to the 33 alarm and 50 nonalarm patterns in the monitoring set. The optimal wavenumber−latent variable combination was 4900−4200 cm−1 and 10 LVs based on the minimum percentage of missed and false alarms (overall 92.8% correct classification). Visual inspection of the 10 PLS loading weights and spectral loadings revealed little contamination by noise-like features, thus helping to confirm that the LVs were extracting nonrandom spectral information. For application to the prediction set, the training and monitoring sets were then combined to form an overall calibration set, which consisted of 247 single-beam spectra. The reference concentrations for the prediction set, alarm threshold value, and the critical concentration were 7.72, 3.00, and −4.72 mM, respectively. A total of 30 381 differential spectra were generated from the calibration data, and the PLS E

DOI: 10.1021/acs.analchem.8b03437 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

Figure 5. (A) Prediction glucose concentration profile for rat B. There were 34 alarm patterns and 165 nonalarm patterns. The alarm threshold of 3.0 mM is shown by the horizontal line. (B) Discriminant score plot for the prediction set with three replicate classifiers. The dashed line indicates a discriminant score of 0.0 that corresponds to the boundary between the alarm and nonalarm data classes. Red, blue, and green symbols denote the discriminant scores for the three replicate classification models. Dots, squares, and circles correspond to correct classifications, missed alarms, and false alarms, respectively.

used with rat A because of the small number of spectra below 3.0 mM in the monitoring set. There were 9601 alarm and 6152 nonalarm patterns in the training set. For each combination of spectral range and PLS LVs, three replicate classifiers were computed, each based on two discriminant functions. A third discriminant was not used here, because it did not separate a significant number of patterns in the training set. The optimal parameters were a spectral range of 4850−4250 cm−1 with 14 LVs. This combination produced 83.7% correct classifications for the 26 alarm and 23 nonalarm spectra in the monitoring set. The higher number of LV’s chosen for rat B is most likely a reflection of the greater variation in these spectra when compared to rat A. The training set and monitoring set were then combined to build a calibration set, which consisted of 227 averaged singlebeam spectra. A total of 25 651 differential spectra were produced and used to compute the PLS scores that formed the patterns. The alarm algorithm was then implemented with the prediction set. The PLS scores computed with the optimal wavenumber−LV combination were partitioned into 8008 alarm and 17 643 nonalarm patterns. The reference concentration, alarm threshold value, and the critical concentration for the prediction set were 7.21, 3.00, and −4.21 mM, respectively. Three replicate classifiers were computed with the calibration PLS patterns. The prediction set contained 34 alarm concentrations and 165 nonalarm concentrations. Figure 5A shows the prediction glucose concentration profile for rat B. The prediction results for the three replicate discriminants were 4, 6, and 6 missed alarms and 29, 16, and 14 false alarms. On the basis of the alarm decision rule described previously, there were 6 missed alarms out of 34 alarm patterns and 16 false alarms out of 165 nonalarm patterns. This corresponded to the successful detection of 82.4% of the alarm events and a false alarm rate of 9.7%. The overall correct classification percentage was 88.9%. This performance is worse than that observed previously with rat A. The discriminant score plot for the prediction set is shown in Figure 5B. The dependence of the discriminant scores on

transient study for hypoglycemia was performed on the same day. Spectral data for rat B were collected over 8.2 h on the basis of 32 averaged scans. Measured glucose concentrations ranged from 2.3 to 13.5 mM (41 to 244 mg/dL). To match the work performed with rat A, every four consecutive spectra were averaged. Spectral diagnostics were computed as described previously for rat A. The mean RMS noise value was 100.2 μAU, and the corresponding median was 74.8 μAU. These values are approximately three times higher than the values obtained for rat A and correspond to less optical throughput for the tissue sampled with rat B. Using the pure-component fitting procedure described previously, the mean water and collagen path lengths for the spectra for rat B were 0.927 ± 0.003 and 0.778 ± 0.004 mm, respectively. By use of PCA, the first seven signal-averaged spectra were judged to be outliers and were removed from the subsequent analysis. As described previously, a reference concentration was assigned to each of the spectra by interpolation of the reference measurements, followed by investigation of the lag time. The optimal lag time was found to be 7 min, and the final glucose concentrations were assigned to each single-beam spectrum on the basis of this time. The glucose concentration profile was then partitioned into a training set, a monitoring set, and a prediction set as shown in Figure 3B. The training set, monitoring set, and the prediction set contained 178, 49, and 199 signal-averaged single-beam spectra, respectively. The training set yielded 15 753 differential spectra. As described previously, a grid search analysis based on cross-validation was performed on the differential spectra to identify the best possible wavenumber and latent variable combinations for further study. The top four wavenumber ranges were 4900−4200, 4900−4300, 4850−4300, and 4850−4250 cm−1. For each range, values of SECV declined steadily, leveling off in the range of 0.5 mM. These spectral ranges and LVs 13−16 were then used to build classification models with PLDA for implementation of the alarm algorithm with the monitoring set. The reference, critical, and alarm threshold concentrations for the monitoring set were 5.48, −1.08, and 4.40 mM, respectively. The alarm threshold was set higher in this case than the value of 3.0 mM F

DOI: 10.1021/acs.analchem.8b03437 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

protocol used here, because the surgery to insert the venous and arterial catheters was judged by the institutional animal care and use committee to be nonsurvivable. Subsequent work will focus on an alternate protocol that will allow multiple day experiments to be performed.

the concentrations can also be observed for rat B. As observed previously with rat A, the majority of the missed and false alarms occurred in the region near the alarm threshold.

■

CONCLUSIONS In this paper, a nocturnal hypoglycemic alarm algorithm was tested with in vivo measurements performed with a rat animal model. By use of glucose clamp procedures, in which glucose or insulin were administered to the animal, data were collected that simulated the glucose excursions that might occur in a human with diabetes during sleep. From these studies, the potential of the alarm algorithm was established. For both rats A and B, the nocturnal alarm algorithm performed well when applied to the prediction data. On average, 87.5% of alarm events were correctly detected, and the occurrence of false alarms was 7.2%. There was clear evidence of spectral drift during the time course of the experiment, but the alarm algorithm was able to overcome this variation without significant negative effects. Missed and false alarms were always in the concentration region near the alarm threshold. The lack of random false alarms provides strong evidence that the alarm algorithm could perform reliably in signaling a hypoglycemic event. Several overall conclusions can be drawn from this work. First, classification performance is negatively impacted by increased spectral noise levels. The issue is a challenging one from an experimental standpoint, however. Increasing the pressure on the tissue interface will result in greater compression of the tissue and a corresponding higher light throughput and lower noise, but such increased pressure will also cause greater variation in the tissue background with time. The ultimate solution would be to have greater incident source power, but that would require a higher brightness light source than is currently available. A higher brightness source might also permit an interface based on a reflectance geometry. This might help to alleviate deformation of the tissue over the time course of the measurements. The issue of concentration lag was addressed empirically in this work by taking a block of data with measured glucose concentrations and fitting it to a quantitative model for concentration as the lag was adjusted. It was hypothesized that the best concentration fit would signal the most appropriate lag. Across the two rats employed, lag values of 11 and 7 min were obtained. This assumption of a constant lag seems justifiable during sleep but is unproven. These empirically derived lag times fall within the ranges reported in the literature.37−39 Additional issues encountered in the in vivo work were a need to detect when the rat had equilibrated with the interface as well as a need to determine when or if the reference spectrum and concentration used to initiate the alarm algorithm required updating. Addressing these questions is a focus of current work. A key component of the work described here was the availability of a database of reference spectra and associated reference glucose concentrations collected from the same rat used in the implementation of the alarm algorithm. This underscores that the noninvasive measurement is tailored to the individual rat. Our experience suggests that the inherent heterogeneity of tissue and variations in physiology preclude calibration of the spectroscopic measurements across rats. Future work will focus on measurements made with the same rat over multiple days. This was not possible with the

■

ASSOCIATED CONTENT

S Supporting Information *

The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acs.analchem.8b03437. Description of surgical procedures and data collection; spectral diagnostic plots for rat A (PDF)

■

AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]; Fax: +1-319-353-1115; Tel: +1-319-335-3214. ORCID

Gary W. Small: 0000-0001-5588-0445 Present Address †

Sanjeewa R. Karunathilaka, Food and Drug Administration, Center for Food Safety and Applied Nutrition, 5100 Paint Branch Pky, College Park, MD 20740, United States Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS The work described in this paper is based on the dissertation of one of the coauthors (S.R.K.). Terry Graham is acknowledged for her contributions in collecting the data and developing the animal care protocol.

■

REFERENCES

(1) Liu, J.; Liu, R.; Xu, K. Appl. Spectrosc. 2015, 69 (11), 1313− 1318. (2) Haxha, S.; Jhoja, J. IEEE Photonics J. 2016, 8 (6), 6805911. (3) Shih, W. C.; Bechtel, K. L.; Rebec, M. V. J. Biomed. Opt. 2015, 20 (5), 051036. (4) Pandey, R.; Paidi, S. K.; Valdez, T. A.; Zhang, C.; Spegazzini, N.; Dasari, R. R.; Barman, I. Acc. Chem. Res. 2017, 50 (2), 264−272. (5) Mesch, M.; Zhang, C.; Braun, P. V.; Giessen, H. ACS Photonics 2015, 2 (4), 475−480. (6) Bandodkar, A. J.; Jia, W.; Yardımcı, C.; Wang, X.; Ramirez, J.; Wang, J. Anal. Chem. 2015, 87 (1), 394−398. (7) Arnold, M. A. Curr. Opin. Biotechnol. 1996, 7, 46−49. (8) Arnold, M. A.; Small, G. W. Anal. Chem. 1990, 62, 1457−1464. (9) Hazen, K. H.; Arnold, M. A.; Small, G. W. Anal. Chim. Acta 1998, 371, 255−267. (10) Zhang, L.; Small, G. W.; Arnold, M. A. Anal. Chem. 2002, 74, 4097−4108. (11) Marquardt, L. A.; Arnold, M. A.; Small, G. W. Anal. Chem. 1993, 65, 3271−3278. (12) Wan, B.; Small, G. W. Anal. Chim. Acta 2010, 681, 63−70. (13) Amerov, A. K.; Chen, J.; Small, G. W.; Arnold, M. A. Anal. Chem. 2005, 77, 4587−4594. (14) Chen, J.; Arnold, M. A.; Small, G. W. Anal. Chem. 2004, 76 (18), 5405−5413. (15) Arnold, M. A.; Burmeister, J. J.; Small, G. W. Anal. Chem. 1998, 70, 1773−1781. (16) Pan, S.; Chung, H.; Arnold, M. A.; Small, G. W. Anal. Chem. 1996, 68, 1124−1135. (17) Bai, C.; Graham, T. L.; Arnold, M. A. Anal. Lett. 2008, 41, 2773−2793. G

DOI: 10.1021/acs.analchem.8b03437 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry (18) Klonoff, D. C.; Ahn, D.; Drincic, A. Diabetes Res. Clin. Pract. 2017, 133, 178−192. (19) Dexcom. Dexcom G6 Continuous Glucose Monitoring System User Guide. https://s3-us-west-2.amazonaws.com/dexcompdf/G6-CGMUsers-Guide.pdf?_ga=2.266989322.588965287.15426588381811654523.1542658838. (20) Buckingham, B. A.; Raghinaru, D.; Cameron, F.; Bequette, B. W.; Chase, H. P.; Maahs, D. M.; Slover, R.; Wadwa, R. P.; Wilson, D. M.; Ly, T.; Aye, T.; Hramiak, I.; Clarson, C.; Stein, R.; Gallego, P. H.; Lum, J.; Sibayan, J.; Kollman, C.; Beck, R. W. Diabetes Care 2015, 38 (7), 1197−1204. (21) Maahs, D. M.; Calhoun, P.; Buckingham, B. A.; Chase, H. P.; Hramiak, I.; Lum, J.; Cameron, F.; Bequette, B. W.; Aye, T.; Paul, T.; Slover, R.; Wadwa, R. P.; Wilson, D. M.; Kollman, C.; Beck, R. W. Diabetes Care 2014, 37 (7), 1885−1891. (22) Karunathilaka, S. R.; Small, G. W. RSC Adv. 2014, 4, 35405− 35414. (23) Karunathilaka, S. R.; Small, G. W. Analyst 2015, 140, 3981− 3988. (24) Karunathilaka, S. R.; Small, G. W. Anal. Chim. Acta 2017, 987, 81−90. (25) Pathirajage, S. R. K. R. Nocturnal Hypoglycemic Alarm Based on Near-Infrared Spectroscopy. Ph.D. Dissertation, University of Iowa, Iowa City, IA, 2014. (26) Libnau, F. O.; Kvalheim, O. M.; Christy, A. A.; Toft, J. Vib. Spectrosc. 1994, 7, 243−254. (27) Olesberg, J. T.; Liu, L.; Van Zee, V.; Arnold, M. A. Anal. Chem. 2006, 78 (1), 215−223. (28) Kaltenbach, T. F.; Small, G. W. Anal. Chem. 1991, 63 (9), 936− 944. (29) Shaffer, R. E.; Small, G. W. Chemom. Intell. Lab. Syst. 1996, 32, 95−109. (30) Haaland, D. M.; Thomas, E. V. Anal. Chem. 1988, 60, 1193− 1202. (31) Jollife, I. T. Principal Component Analysis, 1st ed.; SpringerVerlag: New York, 1986. (32) De Maesschalck, R.; Jouan-Rimbaud, D.; Massart, D. L. Chemom. Intell. Lab. Syst. 2000, 50, 1−18. (33) Fraser, D. M., Ed. Biosensors in the body: continuous in vivo monitoring, 1st ed.; John Wiley & Sons: Chichester, UK, 1997. (34) Schoonen, A. J. M.; Wientjes, K. J. C. Diabetes Technol. Ther. 2003, 5, 589−598. (35) Arnold, M. A.; Olesberg, J. T.; Small, G. W. Near-Infrared Spectroscopy for Noninvasive Glucose Sensing. In In Vivo Glucose Sensing; Cunningham, D. D., Stenken, J. A., Eds.; John Wiley & Sons: Hoboken, NJ, 2009; pp 357−390. (36) Wilson, G. S.; Gifford, R. Biosens. Bioelectron. 2005, 20, 2388− 2403. (37) Shi, T.; Li, D.; Li, G.; Zhang, Y.; Xu, K.; Lu, L. J. Diabetes Res. 2016, 2016, 4596316. (38) Basu, A.; Dube, S.; Veettil, S.; Slama, M.; Kudva, Y. C.; Peyser, T.; Carter, R. E.; Cobelli, C.; Basu, R. J. Diabetes Sci. Technol. 2015, 9 (1), 63−68. (39) Garg, S. K.; Voelmle, M.; Gottlieb, P. A. Diabetes Res. Clin. Pract. 2010, 87, 348−353.

H

DOI: 10.1021/acs.analchem.8b03437 Anal. Chem. XXXX, XXX, XXX−XXX