Proof of Concept for a New Raman-Based Prototype for Noninvasive Glucose Monitoring

Abstract

Background:

Noninvasive glucose monitoring (NIGM) in diabetes is a long-sought-for technology. Among the many attempts Raman spectroscopy was considered as the most promising because of its glucose specificity. In this study, a recently developed prototype (GlucoBeam, RSP Systems A/S, Denmark) was tested in patients with type 1 diabetes to establish calibration models and to demonstrate proof of concept for this device in real use.

Methods:

The NIGM table-top prototype was used by 15 adult subjects with type 1 diabetes for up to 25 days at home and in an in-clinic setting. On each day, the subjects performed at least six measurement units throughout the day. Each measurement unit comprised two capillary blood glucose measurements, two scans with an intermittent scanning continuous glucose monitoring (CGM) system, and two NIGM measurements using the thenar of the subject’s right hand.

Results:

Calibration models were established using data from 19 to 24 days. The remaining 3-8 days were used for independent validation. The mean absolute relative difference of the NIGM prototype was 23.6% ± 13.1% for the outpatient days, 28.2% ± 9.9% for the in-clinic day, and 26.3% ± 10.8% for the complete study. Consensus error grid analysis of the NIGM prototype for the complete study showed 93.6% of values in clinically acceptable zones A and B.

Conclusions:

This proof of concept study demonstrated a practical realization of a Raman-based NIGM device, with performance on par with early-generation CGM systems. The findings will assist in further performance improvements of the device.

Introduction

The International Diabetes Federation reported that 463 million adults worldwide were suffering from diabetes in 2019.¹ In 2045, this number is estimated to increase by approximately 50% to 700 million people. In type 1 and type 2 diabetes, especially for people with an intensified insulin therapy with either multiple daily insulin injections (MDI) or continuous subcutaneous insulin infusion (CSII), self-monitoring of blood glucose is recommended at least 4 times per day for an adequate therapeutic treatment,² but testing capillary blood can be painful and time consuming, sometimes resulting in poor compliance.³ In the past decades, minimally invasive and noninvasive glucose monitoring (NIGM) has gained considerable attention.⁴ Continuous glucose monitoring (CGM) with indwelling sensors has focused on the interstitial compartment as an alternative for capillary testing. Despite the increasing number of CGM users, CGM is still quite invasive and not free of complications. In addition, most systems require purchase and insertion of sensors in intervals of 6-14 days.⁵

In contrast, noninvasive technologies, including electrical, thermal, acoustical, and optical methodologies, although desired for three decades, could not yet achieve a place in clinical routine.⁴ Among the optical transdermal methods, Raman spectroscopy has gained considerable interest because of its glucose specificity.⁶ Raman spectroscopy is based on the inelastic scattering of light that interacts with a molecule. In this case, laser light is absorbed by the molecule (eg, glucose), leading to a temporarily excited energetic state of the molecule. In conventional (ie, spontaneous) in vivo Raman spectroscopy, an impinging monochromatic laser beam interacts with the tissue of interest, where the constituting molecules (eg, glucose) may become excited to higher vibrational modes by the inelastic scattering of photons, thereby leading to a concurrent shift of frequency of the scattered light. By recording the scattered intensity as a function of change in frequency, the resulting spectrum can be used for quantitative analysis, which, for example, allows determination of the glucose concentration in the tissue.

In the past, several attempts have been made to demonstrate the feasibility of Raman spectroscopy in glucose monitoring;^7-9 however, it has proven impossible to solve the intractable problem concerning stability of calibration, both in terms of 1) stability in time and 2) stability when changing probe positioning on the skin. This problem needs to be solved to realize a practical, useful device, which was recently achieved with the GlucoBeam device¹⁰ (RSP Systems A/S, Denmark), which was also used in this study. Moreover, a recent study by Kang et al. has demonstrated that the glucose fingerprints from in vivo tissue spectra are directly observable by carefully controlling influencing factors.⁶ These findings showed that Raman spectroscopy can in principle be used to measure glucose in vivo directly.

In the present investigation, a novel table-top prototype¹⁰ of an NIGM device was tested in a group of insulin-treated subjects with type 1 diabetes with the aim to assess calibration models and performance in comparison with an intermittent scanning CGM (iscCGM) system,¹¹ as well as to evaluate the feasibility of the method in uncontrolled, real-life use situations. The findings are aimed to assist in the further performance improvements of the device.

Methods

This was an open, prospective, bicenter, two-arm study. Because study design and use of NIGM device were different between the two study arms, this report focuses only on the results of study arm 2, which was performed between November 2018 and March 2019 at the Institut für Diabetes-Technologie, Forschungs- und Entwicklungsgesellschaft mbH an der Universität Ulm (IfDT), Germany, under consideration of the Declaration of Helsinki and in compliance with the Guideline for Good Clinical Practice and the national regulations and provisions. The study protocol was approved by the responsible Ethics Committee and exempted from approval by the German Federal Institute for Drugs and Medical Devices. The study was registered at clinicaltrials.gov (NCT03781232).

Participants

Adult subjects with type 1 or type 2 diabetes and a skin phototype of I to IV (measured with a commercial skin tone sensor; DEESS Demi II GP531, Shenzhen GSD Tech Co., Ltd., China) were eligible for the study. Skin phototype depends on the amount of melanin in the skin and is divided into six categories, which reflect the skin’s reaction to UV light (skin phototype I: always burns, never tans; skin phototype IV: burns minimally, always tans well).¹² After providing written informed consent, subjects were screened for potential enrolment. Exclusion criteria were severe hypoglycemia in the past 3 months; hypoglycemia unawareness; pregnancy or lactation period; known severe allergy to medical-grade adhesive or isopropyl alcohol (used to clean the skin); inability to comply with the study procedures (due to, eg, known psychiatric diagnoses, lack of cognitive ability, alcohol dependency, drug use, or psychosocial overload); severe diabetes-related complications (eg, advanced autonomic neuropathy, kidney disease, foot ulcers, legal blindness, or symptomatic cardiovascular disease as evidenced by a history of cardiovascular episode(s)); systemic or topical administration of glucocorticoids for the past 7 days; inability to hold arm or hand still (including tremors and Parkinson’s disease); intake of salicylic acid or higher doses of ascorbic acid; or undergoing dialysis treatment. After screening, 15 subjects with type 1 diabetes, 2 males and 13 females, were included in the study. Their mean age was 56 years. Of the 15 subjects, 11 used MDI, whereas 4 were on CSII therapy. Study duration for each subject was 28 calendar days.

Study Devices

The prototype NIGM device (GlucoBeam, RSP Systems A/S, Denmark) was a portable, stand-alone device, including a confocal Raman spectrometer configured for NIGM in people with diabetes. The device made use of critical-depth Raman spectroscopy, such that the Raman signal is spatially filtered to select signals from about 300 µm in the thenar tissue, which may include vascular, interstitial, and intracellular compartments. The NIGM uses a confocal Raman setup in which the illumination is strongly focused to a spot with width and depth of ~40 and ~250 µm, respectively. The spot size is measured in air and corresponds to full-width-at-half-maximum values. One measurement unit of the NIGM device comprised two measurements with measurement durations of 2 and 1 minute, respectively, within which multiple spectra were recorded. The integration time was subject-dependent and set to properly use the dynamic range of the image sensor. With an output power of approximately 250 mW, integration times varied between 1.5 and 2.5 seconds. The spectral range is 300-1615 cm⁻¹, and within this range the average resolution is approximately 8 cm⁻¹. The NIGM uses a laser wavelength of 830 nm. The signal in the signal-to-noise ratio, which is the glucose spectrum in this case, is not readily tractable, because presence of glucose can only be identified using advanced spectral preprocessing and multivariate data analysis.¹³ Further details, such as technical characteristics of the prototype and first results in subjects with diabetes, were described recently by Lundsgaard-Nielsen et al.¹⁰ The device was still in the explorative phase, and this article reports on the first results obtained from a combined outpatient and in-clinic trial with prototype P0.3.

The FreeStyle Libre (Abbott Diabetes Care, Alameda, CA, USA) was used as a system for iscCGM (flash glucose monitoring). The version used in the study could be worn for up to 14 days and it was factory-calibrated. It measured glucose levels every minute in the interstitial fluid and continuously stored one value every 15 minutes. The system needed to be actively scanned to obtain current glucose information, ie, values and glucose traces, to be displayed on the device’s reader. Scans had to be performed at least every 8 hours to retain the whole daily glycemic data.¹⁴

Comparison Measurements

Capillary comparison measurements, used to compare values of the CGM system and the NIGM prototype during home and in-clinic session days, were performed using the Contour next ONE (Ascensia Diabetes Care GmbH, Basel, Switzerland) blood glucose monitoring system (BGMS). Accuracy of the glucose test strip batch that was used in the study was verified by performing additional, twice-daily comparison measurements with a hexokinase-based laboratory glucose analyzer (Cobas Integra 400 plus; Roche Instrument Center, Rotkreuz, Switzerland) during in-clinic session days. The conformity to traceability requirements of the method to ISO 17511¹⁵ was assured by the analyzer’s manufacturer and verified by higher-order controls.

Study Design and Procedures

After inclusion, subjects were trained on all study devices by study staff at their first study site visit. Because the study duration was 28 calendar days and the sensor lifetime of the iscCGM system is 14 days, each subject used two sensors during the study; replacement of sensors was performed by the subjects themselves. On each outpatient study day, the subjects performed at least six measurement units distributed throughout the day. Each measurement unit comprised two capillary BGMS measurements, two scans of the iscCGM, and two NIGM measurements using the thenar of the right hand of the subject. Glucose measurements were performed in the following order: BGMS—iscCGM—NIGM—BGMS—iscCGM—NIGM. Except for two in-clinic “glucose excursion days” (GEDs), one on day 5 or 6 and one on day 27, the subjects treated their diabetes in the usual manner and operated the NIGM device at home without guidance from trained personnel. Rapid glucose changes were induced on GEDs by eating a breakfast with a high glycemic index after an overnight fasting period and delivering a delayed and increased insulin bolus. To increase the variability between the glucose curves of the subjects, the insulin bolus was administered either 30 and 60 minutes or 30, 60, and 90 minutes after start of meal intake and with different bolus sizes (50% for two bolus injections or 50%/25%/25% for three bolus injections). After these meals, subjects performed measurement units every 15 minutes for the following 6-7 hours. Subjects were closely monitored by a study physician during these induced rapid glucose changes. On each of these GEDs, capillary glucose samples were obtained for subsequent measurements on a hexokinase-based laboratory analyzer (Cobas Integra 400 plus, Roche Instrument Center, Switzerland).

Data Analysis

For the NIGM prototype, one glucose concentration was available per measurement unit, ie, both individual measurements contributed to one result. Therefore, average values were calculated for the two BGMS measurements and the two iscCGM measurements, respectively, in each measurement unit to account for potential systematic differences between the individual measurements due to glucose concentration changes. Analyses were based on time points for which each of the three values were available. Data from the first 19 to 24 days were used by the manufacturer for calibration of the NIGM system. Calibration models are based on partial least squares (PLS) regression, where the individual regression models were built using data from the first 24, 19, and 22 days for subjects enrolled in rounds 1, 2, and 3, respectively. The varying number of calibration days was part of the explorative design to investigate how different amounts of calibration data affect the measurement accuracy. During data preparation, all spectra captured during one measurement unit were averaged to a single spectrum, while the associated two BGM measurements were averaged to a single reference value. With the stated number of calibration days, PLS models are based on 120-150 paired spectra and reference values. Before model building, the spectra were further processed by Savitzky–Golay smoothing (seven-point window, first-order polynomial) and corrected for varying intensities and fluorescence backgrounds by 1st-order extended multiplicative scatter correction.¹⁶ Spectral outliers were identified and removed by calculation of Q-residual and Hotelling’s T2 statistics, requiring that spectra (to be included in the calibration model) lie within the 99% confidence interval.¹⁷ Finally, the spectra and reference values were mean-centered and supplied to the PLS algorithm. The number of PLS components were determined from minimization of the root mean squared error of 10-fold, contiguous cross-validation. The data from the remaining days (“complete study”), including one GED for each subject, were used for independent validation of the calibration. The last GED (day 27) was blinded and evaluated by IfDT in order to ensure that independent validation was rigorously observed and the data were independently assessed.

For data from validation days, mean absolute relative difference (MARD) was calculated for the iscCGM device and the NIGM prototype in comparison with BGMS results for each subject, and consensus error grid (CEG) analysis was performed for the prototype. In addition, the relative bias (%) was calculated according to Bland and Altman¹⁸ by using the following formula:

$$\frac{1}{n}\sum ^{n}2\times \frac{\left(GM-comparison\right)}{\left(GM+comparison\right)}$$

in which GM (glucose monitoring) is a single NIGM or mean iscCGM measurement result, comparison is the mean result of the comparison measurement (BGMS), and n is the number of all GM measurement results.

In addition, time in range (70-180 mg/dL), time below range (<70 mg/dL), and time above range (>180 mg/dL) were calculated by dividing the number of values in the specific glucose range by the total number of glucose values.

In a supplemental analysis, measurement accuracy of the BGMS and the iscCGM system was assessed in comparison with the hexokinase-based laboratory analyzer for the two GEDs. This analysis was not done for the NIGM system, because GED 1 was part of the calibration period.

Furthermore, for each subject’s calibration model, an analysis of the standard error of prediction (SEP) for all validation points and for the blinded GED 2, respectively, was performed.^19,20

Results

Calibration Models

For each subject separately, a calibration model was established. Calibration was based on data from 24 days for the first group of 5 subjects, 19 days for the second group of 5 subjects, and 22 days for the third group of 5 subjects, to investigate the effect of different amounts of calibration data. Calibration data were collected contiguously prior to the validation data.

Accuracy

For the BGMS, which was used as comparison method and for NIGM calibration, 95.0% of the results on GEDs 1 and 2 were within ±15 mg/dL of the hexokinase measurements at glucose concentrations <100 mg/dL or within ±15% at glucose concentrations ≥100 mg/dL. A mean relative bias of 5.3% ± 1.5% was observed (Supplemental Table 1).

Based on data from the validation days, the iscCGM system showed an MARD of 15.6% ± 6.4% for the outpatient days, 22.7% ± 9.2% for the in-clinic day, and 18.8% ± 5.8% for the complete study. In comparison, the MARD of the NIGM prototype was 23.6% ± 13.1% for the outpatient days, 28.2% ± 9.9% for the blinded in-clinic day, and 26.3% ± 10.8% for the complete study. Aggregated MARD results were similar (Table 1).

Table 1.

Subject-Specific Accuracy Results of the iscCGM System and the NIGM Prototype for the Outpatient Days, the In-Clinic Day, and the Complete Validation Phase.

Subject #	iscCGM system									NIGM prototype
	Outpatient days			In-clinic day (with glucose excursion)			Complete study			Outpatient days			In-clinic day (with glucose excursion)			Complete study
	MARD (%)	Relative bias (%)	n	MARD (%)	Relative bias (%)	n	MARD (%)	Relative bias (%)	n	MARD (%)	Relative bias (%)	n	MARD (%)	Relative bias (%)	n	MARD (%)	Relative bias (%)	n
1	18.3	−20.1	10	17.9	−20.3	15	18.0	−20.2	25	9.1	−8.9	10	31.3	−32.2	15	22.4	−22.9	25
2	18.6	−20.8	15	27.2	−31.6	14	22.7	−26.0	29	26.9	9.6	15	23.7	16.7	14	25.4	13.1	29
3	12.1	−6.1	11	13.6	−6.0	20	13.1	−6.0	31	19.2	−0.1	11	35.7	11.1	20	29.9	7.1	31
4	8.0	−8.4	7	14.0	−12.8	22	12.5	−11.8	29	19.0	−6.4	7	30.0	−15.8	22	27.3	−13.5	29
5	14.9	−15.5	29	20.5	−23.0	17	17.0	−18.3	46	15.5	−10.6	29	21.4	−1.0	17	17.7	−7.1	46
6	11.1	−10.8	27	45.3	−59.6	11	21.0	−24.9	38	18.0	−5.9	27	21.0	10.1	11	18.9	−1.3	38
7	26.6	−30.8	9	27.5	−32.2	27	27.3	−31.8	36	20.8	9.8	9	24.2	−7.3	27	23.4	−3.0	36
8	11.9	−11.9	22	24.7	−29.2	8	15.3	−16.5	30	22.1	−8.2	22	19.5	−10.8	8	21.4	−8.9	30
9	24.1	−29.1	28	20.8	−23.4	21	22.7	−26.7	49	23.0	1.7	28	24.4	−29.1	21	23.6	−11.5	49
10	14.8	−16.2	14	21.1	−23.8	15	18.1	−20.1	29	12.7	3.0	14	24.9	−29.2	15	19.0	−13.6	29
11	12.9	0.9	29	7.6	0.9	13	11.3	0.9	42	19.4	2.1	29	27.7	11.6	13	22.0	5.1	42
12	11.7	−12.5	16	19.8	−22.5	25	16.6	−18.6	41	28.6	18.0	16	24.0	0.5	25	25.8	7.3	41
13	28.7	−33.9	23	35.5	−43.7	28	32.5	−39.3	51	29.2	4.2	23	34.1	−5.7	28	31.9	−1.2	51
14	8.7	−1.9	20	19.1	−17.9	23	14.3	−10.4	43	24.8	3.0	20	21.4	−0.7	23	23.0	1.0	43
15	12.4	−11.3	20	25.9	−29.9	26	20.0	−21.8	46	66.3	44.0	20	59.5	33.6	26	62.5	38.1	46
Range	8.0-28.7	−33.9 to 0.9		7.6-45.3	−59.6 to 0.9		11.3-32.5	−39.3 to 0.9		9.1-66.3	−10.6 to 44.0		19.5-59.5	−32.2 to 33.6		17.7-62.5	−22.9 to 38.1
Mean	15.6	−15.2		22.7	−25.0		18.8	−19.4		23.6	3.7		28.2	−3.2		26.3	−0.8
Median	12.9	−12.5		20.8	−23.4		18.0	−20.1		20.8	2.1		24.4	−1.0		23.4	−1.3
SD	6.4	10.2		9.2	14.5		5.8	10.1		13.1	13.7		9.9	18.5		10.8	14.4
Aggregated mean	15.7	−14.9		22.7	−24.9		19.3	−20.0		24.2	3.3		29.6	−2.6		26.9	0.3
Aggregated median	14.6	−15.3		22.8	−25.5		17.8	−19.0		17.5	−0.1		23.0	−4.8		19.6	−2.1
Aggregated SD	10.7	15.9		12.1	18.4		11.9	17.9		25.5	28.1		29.2	35.1		27.5	31.9
N			280			285			565			280			285			565

MARD and relative bias (n = 15) were calculated against BGMS values (data from validation days).

BGMS, blood glucose monitoring system; iscCGM, intermittent scanning continuous glucose monitoring; MARD, mean absolute relative difference; NIGM, noninvasive glucose monitoring.

The mean relative bias of the iscCGM system compared to the BGMS was –15.2% ± 10.2%, –25.0% ± 14.5%, and –19.4% ± 10.1% for the outpatient days, the in-clinic day, and the complete study, respectively. In contrast to that, the NIGM prototype had a mean relative bias of 3.7% ± 13.7% for the outpatient days, –3.2% ± 18.5% for the in-clinic day, and –0.8% ± 14.4% for the complete study (Table 1).

CEG analysis of the NIGM prototype showed 58.6% and 36.4% of values in clinically acceptable zones A and B for the outpatient days, 48.1% and 44.2% for the in-clinic day, and 53.3% and 40.4% for the complete study, respectively. The remaining values fell within zones C (4.6%, 7.0%, and 5.8%, respectively) and D (0.4%, 0.7%, and 0.5%, respectively). No values were found in zone E (Figure 1a-c).

Figure 1.

CEG distribution of paired points of (a) outpatient performance, (b) in-clinic performance, and (c) overall performance. All paired points correspond to results obtained during validation phase.

CEG, consensus error grid.

Calculation of the SEP of the NIGM prototype from each subject’s calibration model showed comparable results between all validation points (48.3 ± 7.8) and the GED 2 (52.6 ± 11.2) (Supplemental Table 2).

Figure 2 and Supplemental Table 3 show the mean glucose kinetics of the three systems (BGMS, iscCGM system, and NIGM prototype) after the breakfast meal challenge. There was a good match between NIGM and BGMS data, but iscCGM data were considerably lower, similar to previously reported findings.^21-23 In general, the match between the capillary BGMS and the two interstitial systems improved with lower rates of change. During the initial 60 minutes after the standardized meal, the rates of change were as follows: BGMS, 1.6 ± 2.0 mg/dL/min, iscCGM, 1.3 ± 1.4 mg/dL/min, and NIGM, 0.9 ± 1.6 mg/dL/min (mean ± SD).

Figure 2.

Mean glucose values (±SD) for BGMS, iscCGM system, and the NIGM prototype in a period of 30 minutes prior to and 450 minutes after the intake of a carbohydrate-rich breakfast. Subjects were fasting overnight. Each data point is based on time points for which each of the three values were available (n = 2 to n = 15). Orange arrows indicate time points for insulin delivery (30, 60, or 90 minutes after start of meal intake).

BGMS, blood glucose monitoring system; iscCGM, intermittent scanning continuous glucose monitoring; NIGM, noninvasive glucose monitoring.

The calculated time in range shows high similarity between the NIGM prototype and the BGMS for the times spent in, below, or above target range (Figure 3). In contrast, the time below range was higher and time above range was lower for the iscCGM system than for the other systems.

Figure 3.

Time in range during the meal challenge for BGMS, iscCGM system, and the NIGM prototype.

BGMS, blood glucose monitoring system; GED, glucose excursion day; iscCGM, intermittent scanning continuous glucose monitoring; NIGM, noninvasive glucose monitoring.

Discussion

In the present study, glucose concentration data were collected to establish calibration models of a newly developed NIGM device, as well as to show proof of concept of the technical and practical realization under uncontrolled daily-life conditions. Calibration models could be established with BGMS data from 19 to 24 days. Future versions of the NIGM prototype could likely provide adequate calibration with less data, as a database with paired thenar spectra and reference measurements could guide and improve the calibration models, such that less individual device data would be needed. Proof of concept was shown in a cohort of 15 subjects with type 1 diabetes who were expected to exhibit considerable individual variability in glucose excursions and response to a standardized breakfast meal. Accordingly, relatively high MARD values were found for both the new NIGM device and the iscCGM system. The NIGM prototype could follow rapid glucose changes indicating the established calibration works for practical useful physiological glucose measurement, but a delay between the capillary and the interstitial compartment was suggested by the different rates of increasing glucose concentrations during the first 60 minutes after the meal challenge. The small sample size at individual time points during this experiment may also have had an influence on the results.

Yet, in the CEG, 94% of the values were in clinically acceptable zones, and the bias was low when compared to the BGMS. In addition, the calculated times in, above, and below range were comparable for the NIGM device and the BGMS. These data indicate that the calibration models were, in general, suitable. Nevertheless, it should be kept in mind that whole-tissue measurements from the vascular, interstitial, and intracellular compartments of the NIGM device were compared with capillary and interstitial glucose concentration values, respectively. The reported results are the first of their kind from this NIGM prototype P0.3, with independent NIGM measurements being used during the clinical trial. As such, the limited knowledge, particularly in the beginning of the trial, on how to preprocess spectra and generate calibration models leaves room for improvement in future studies or retrospective analyses. However, the main intention of the clinical trial and associated data analysis was to provide proof of a practical realization of a NIGM device.

Interestingly, a considerable negative bias was found for the iscCGM system in comparison with BGMS values. Although some reports about this iscCGM system indicate underestimation of glucose concentrations in comparison to reference measurements,^21-23 the extent was larger than expected. This finding could be influenced, at least to some degree, by the comparison measurement method.

Dingari et al. reported in 2011²⁴ that the prediction of real glucose concentrations with Raman spectroscopy-based data is not possible, because of inadequate assignment of glucose concentration profiles to Raman spectra. However, the fundamental understanding of the underlying mechanisms and technical aspects of Raman-based NIGM devices has matured over the past few years. Some recent studies successfully verified the underlying technique used in this prototype.^10,25 The performance of this prototype is comparable to findings from earlier stages of BGMS and to that of early generations of commercially available CGM systems.^26-28 As shown before with a similar prototype,¹⁰ glucose monitoring was performed equally well under outpatient conditions. The study can be seen as a further step to the goal of a wearable miniaturized version.

Conclusions

This proof of concept study of a practical Raman spectroscopy-based NIGM device showed promising results. Calibration models could be identified, and independent validation data, including a blinded in-clinic day with glucose excursion, demonstrated accuracy comparable to early-generation CGM systems. Overall, this study indicates that Raman spectroscopy and the particular critical-depth implementation feature the prerequisites for practical NIGM, with the current findings representing an important stepping stone toward the development of the final device.