communications medicine Article

A Nature Portfolio journal

[Link]

Clinical validation of noninvasive blood

glucose measurements by midinfrared
spectroscopy
Check for updates
1 1 1 1 1 1
Michael Kaluza , Sergius Janik , Thorsten Lubinski , Mattia Saita , Peter Lachmann , Luca Canini ,
Valentino Lepro 1 & Werner Mäntele 1,2

Abstract Plain language summary

1234567890():,;
1234567890():,;

Background Non-invasive glucose monitoring (NIGM) is considered the holy grail of People with diabetes cannot regulate the
diabetes technology. Currently, diabetes is predominantly managed based on finger amount of a sugar called glucose in their
pricking and analysis of a drop of blood with test strips fitting into a glucometer. This blood. They need to regularly check the
invasive, painful and uncomfortable procedure is one of the reasons for insufficient diabetes amount of glucose in their blood so they can
management. Sensors for continuous glucose monitoring (CGM) in interstitial fluid (ISF) take treatments that adjust the levels. This
using a subcutaneous microscopic filament present a minimally invasive alternative. procedure is painful and uncomfortable as it
Worldwide, there is intense research and development for a true non-invasive glucose involves pricking the skin to make it bleed. We
measurement. used an infrared laser that enables a painless
Methods DiaMonTech has demonstrated a NIGM technology that targets glucose in ISF. An determination of the amount of glucose
infrared beam from a quantum cascade laser excites glucose molecules at wavelengths present in skin. Our device was tested in 36
between 8 and 12 µm, where glucose has specific fingerprint absorbance. Absorption people with and without diabetes and showed
results in a small amount of heat in the skin, which can be detected on the surface using a an accuracy similar to other glucose monitors
photothermal deflection technique. This procedure is painless, harmless, and does not approved by the US food and drug
require consumables. administration (FDA). Our method could make
Results We report here a single-center clinical test with 36 individuals ([Link] ID: blood glucose checks easier and more
NCT06088615). The accuracy of the NIGM device was evaluated in two subsequent comfortable in the future and improve
sessions per individual, with different amounts of calibration data. Four different algorithms diabetes management.
were tested for data analysis. The accuracy for the best algorithm, expressed as Mean
Absolute Relative Difference was 20.7 % and 19.6 % for the two sessions.
Conclusions This is equivalent to the performance of early CGM systems cleared by the FDA
for adjunctive use by people with diabetes. It demonstrates that glucose can be reliably
measured with this non-invasive technology and opens new perspectives for a better
management of diabetes.

Non-invasive glucose monitoring is a complex problem and considered procedure produces problematic waste2, bears the risk of infection, and is
the holy grail of diabetes technology, as it can improve the life of many uncomfortable and painful. Consequently, users of traditional devices do
humans around the globe. An increasing number of people (currently not perform the close-meshed monitoring needed to keep their BG in the
589 million adults) worldwide live with diabetes1. At present, diabetes “green range“ (approx. 70–180 mg/dL). The situation has considerably
cannot be cured, instead it is managed by strict control of blood glucose improved by the introduction of minimally invasive sensors for con-
(BG) and adaptation of food intake, physical activity, medication, and if tinuous glucose monitoring in the interstitial fluid (ISF), where a short
necessary, by insulin administration. BG monitoring is predominantly filament is introduced into the skin using a needle (for a review, see
performed invasively by finger pricking and analysis of a drop of blood ref. 3). These sensors are particularly appreciated by people with type 1
using a glucometer in combination with enzymatic test strips. This diabetes who require close-meshed control with warnings for hypo- and

1
DiaMonTech AG, Berlin, Germany. 2Institut für Biophysik, Johann Wolfgang Goethe-Universität, Frankfurt am Main, Germany.
e-mail: [Link]@[Link]; maentele@[Link]

Communications Medicine | (2025)5:509 1

[Link] Article

hyperglycemia, although skin irritations (the sensor patch is worn for up Optical technologies for NIGM targeting glucose in skin have been
to two weeks) have been reported4. developed based on the absorption of visible, near-infrared (NIR), mid-
In the search for a truly non-invasive solution, numerous technological infrared (MIR), Terahertz (THz) and microwave (MW) radiation as well as
attempts have been made over the past >20 years; however, no reliable on Raman scattering. A glucose solution in water is colorless, i.e., visible light
glucometer function has been demonstrated in a consumer device. Some of does not exhibit any specific absorption for the glucose molecule. Absorp-
these efforts included physiological parameters as proxies that parallel BG, tion of a glucose solution in the ultraviolet region (UV) is not specific, as is its
such as temperature, electric conductivity, impedance of skin, or ultrasound absorbance in the THz or MW range. However, glucose exhibits a distinct
and microwave transmission of skin (for a recent review, see ref. 5). The vibrational signature in the MIR wavelength range from ~8 to 11 μm, arising
specificity of these proxies for BG concentrations, however, is too low to from coupled -C-O- stretching and -O-H bending vibrations of the mole-
allow a safe measurement for people with diabetes. Other attempts use cule that also are the basis of its Raman scattering behavior (Fig. 1). This
optical techniques such as near-infrared spectroscopy (for a recent review, signature is highly specific for the glucose molecule and allows distinction
see ref. 6), Raman spectroscopy7–9, or mid-infrared spectroscopy10–13 directly from other molecules in the skin but requires more sophisticated light
targeting the glucose molecule. sources, optics and detection. As for many molecules, the fundamental
Several approaches included the analysis of easily accessible body fluids vibrations in the MIR are accompanied by overtones (“harmonics”), mul-
such as urine, saliva, sweat and tear fluid as potential proxies for blood. tiples of the fundamental modes that are found in the NIR, in the range
Although these proxy fluids may contain some glucose, its concentration is between approx. 1 to 3 μm. The absorption cross-sections of these overtones
much lower than in blood, it does not exhibit an explicit relation to BG, or it are much weaker (more than 10–100 times) than those of the fundamentals.
exhibits delays with respect to BG that could pose a health risk for people In addition, overtones of all molecules in skin are found in a relatively
with diabetes. narrow spectral region, and furthermore, are affected by temperature in
ISF has turned out to be the most suitable proxy for blood. It is found in position and intensity. In spite of all these drawbacks, the easy availability of
skin layers immediately below the stratum corneum at a depth from 20 to NIR light sources (e.g., Telecom lasers), optics, and detectors have led to
200 μm below the skin surface and thus accessible with many optical many attempts to develop NIR-based glucose sensing.
technologies. It typically contains 85–90% of BG, equilibrated with the The outstanding MIR signature of glucose (Fig. 1) has prompted
capillary blood flow by diffusion. Depending on blood circulation and local DiaMonTech to develop a sensor based on recently developed MIR sources,
capillarization at the respective skin parts, it follows BG with delays of only a called quantum cascade lasers (QCL), and photothermal detection15–17. A
few minutes14. ISF contains essentially the blood protein albumin, glucose, block diagram of the working principle and the signal generation is provided
lactate and ions10. CGM systems that measure with a small filament under in the section “Materials and Methods”.
the skin demonstrate that ISF is a reliable proxy for blood. Briefly, pulsed MIR light from the QCL is directed through an internal
reflection element (IRE) made from IR transparent material into the skin,
where it penetrates up to 100 μm and is absorbed in ISF-containing layers
(Fig. 2a). Glucose in these layers is vibrationally excited from the ground
level v0 to the first vibrational level v1, immediately (within ~10−12 s) fol-
lowed by thermal relaxation. The relaxation results in the deposition of a
small amount of heat around the glucose molecule. Its amount corresponds
to the energy of the absorbed MIR photon.
The heat spreads out to the surface of the skin and enters the IRE, where
it forms a temporary thermal lens, a region where the refractive index is
altered. To probe the thermal lens, a second visible or near-infrared laser
beam is directed into the IRE and passes through the thermal lens, where it is
deflected. The deflection is detected by a position-sensitive photo detector.
The probe beam deflection is directly related to the pump laser power and
the glucose concentration. Variation of the pump laser wavelength in the
spectral region of glucose absorption allows to probe the glucose spectrum
and to determine glucose concentration even on the background of other
molecules in skin.
Based on this technology, DiaMonTech has developed prototypes of a
table top glucometer (“D-Base”) that include a commercial external cavity-
tuned QCL, photothermal sensing unit, data acquisition unit and signal
Fig. 1 | Infrared vibrational spectra of molecules relevant for skin in aqueous model processing hardware and software (Fig. 2b).
solutions: glucose, lactate and albumin.

Fig. 2 | Measurement principle and test device. a

Visualization of photothermal detection of glucose
in skin: In yellow the MIR excitation beam; in red the
probe beam being deflected by the thermal lens. b
D-Base with open display and mechanical lid
(arm rest).

Communications Medicine | (2025)5:509 2

[Link] Article

and it is based on a 2D position-sensitive photodetector from Hamamatsu

(Hamamatsu Photonics KK, Japan).
The raw data processing is handled by a custom-built data acquisition
system (DAQ). The computing unit is based on a Raspberry Pi 4 (Raspberry
Pi Holdings, Cambridge, England) with a 64 bit Linux OS and custom-built
software.
A custom-made ergonomic wrist-rest accessory was designed and 3D
printed. The object is designed to resemble a wristwatch and fastened with a
Velcro strip to the arm. A mechanism allows the wristband to be securely
and firmly clicked to the D-Base when the subject is ready for the mea-
surements. This system ideally allows to measure the same skin spot over the
entire session, improving measurement stability and reproducibility.

Study setup
The study reported here was performed between November 2023 and April
2024 by an independent institute (Institut für Diabetes-Technologie, IfDT)
Fig. 3 | Mid-infrared measurement of glucose in skin. Block diagram of the in Ulm (Germany) as a prospective study with 36 individuals (19 female, 17
Working principle and main components of the D-Base. male) aged between 20 and 75 years. The aim of this study was to assess the
accuracy of this NIGM approach and to collect data for further product
developments. The participants took part in three sessions on three different
We report here a single-center clinical evaluation with 36 individuals. days. The first session was used for calibration purposes only, while the
The accuracy of the NIGM device was evaluated in two subsequent sessions second session served as test session using only the first session for cali-
per individual, with different amounts of calibration data. Four different bration. In the third session, the device was validated using the previous two
algorithms were tested for data analysis. The accuracy for the best algorithm, sessions for calibration.
expressed as Mean Absolute Relative Difference, was 20.7% and 19.6% for The study was planned in the summer of 2023 and the study design
the two sessions. This is equivalent to the performance of early CGM sys- approved by the ethics committee and the competent authority. It was
tems cleared by the FDA for adjunctive use by people with diabetes. It designed in a typical fashion for assessing the accuracy of a glucose mea-
demonstrates that glucose can be reliably measured with this non-invasive surement technology with an observation time of about 5 h per session and
technology and opens new perspectives for a better management of diabetes. including an induced glucose change with a standardized meal and regular
measurements with the reference and the test method. The study was
Materials and methods approved by the Ethics Commission of the Landesärztekammer Baden-
The test device D-Base Württemberg on September 19, 2023, and registered by the BfArM on
The measurement principle has been described earlier16, has essentially October 5, 2023, as well as at [Link] (ID: NCT06088615). It was
remained unchanged and is illustrated in Fig. 3. The device has been conducted by the IfDT, an independent contract research organization
improved, and the main changes are reported here. accredited by the Deutsche Akkreditierungsstelle GmbH (DAkkS) as testing
The D-Base used in the study is the version “D-Base 2.0 prototype B”. laboratory according to DIN EN ISO/IEC 17025 in terms of several test
Central component is a commercially available external cavity tunable procedures according to DIN EN ISO 15197:2015.
quantum cascade laser (EC-QCL) as MID-IR source, the DRS Daylight In total, 36 participants were planned, recruited and enrolled for this
Solutions Hedgehog (Leonardo DRS, San Diego, USA). This QCL is capable study. A signed informed consent to participate in the study, including an
of emitting light in the 8.3–10.5 μm wavelength range and is used to excite understanding and willingness to follow the protocol, was required.
the glucose molecules in the skin. The laser is operated in pulsed mode with Volunteers comprised people with diabetes (13 type-1 diabetes, 17 type-2
100 ns pulses at 10% duty cycle. Additionally, the QCL is controlled by a diabetes) as well as healthy individuals (6). One participant discontinued the
proprietary signal generator board that introduces an additional modula- study during the second session due to illness unrelated to the study, which
tion of the laser, allowing emission only for 50% of the time with a frequency led to 35 participants in total. One participant could only complete two of
of 57 Hz while sweeping through the spectral range. The average output the three sessions due to a medical appointment unrelated to the study,
power of the QCL in this configuration is around 10 mW. This EC-QCL is hence a total number of 105 sessions were completed.
classified as a Class 3R laser. The accessible emission limit (AEL) for the Important inclusion criteria for the test participants were clinical
QCL based on the parameters used is 5000 W/m², while the measured value diagnosis of type-1 diabetes, type-2 diabetes, or no diabetes. For subjects
by an accredited testing site was 1373 W/m² of irradiance. with diabetes HbA1c < 10% was required. All participants had to be at least
The probe laser to read out the thermal lens is used in total internal 18 years old, all genders could be included. All subjects were of white
reflection in the IRE, thus only a few μW of power are accessible to the user. ethnicity.
The AEL for the probe laser is 390 μW, while the measured value was below No selection was made for skin pigmentation. Melanin (responsible for
10 μW. Both lasers are safely protected in the device and precautions have skin pigmentation) is produced by melanocytes in the deepest layer of the
been taken to avoid the beam being viewed directly or through external epidermis (”basal layer”) and stored there in intracellular units called mel-
optics. The device has passed the laser safety tests, which are based on the anosomes. This melanin distribution is applicable to all skin colors. Since the
MPE (Maximum Permissible Exposure, see: IEC 60825-1 regarding laser optical penetration depth of the infrared laser is only about 100 μm, no
safety) for skin. Our device works below 50% of the maximum safety significant contribution from skin pigmentation is expected, which is a
threshold. substantial advantage of our infrared technology.
The optical measurement interface is the so-called internal reflection The inner side of the distal forearm, preferably close to the wrist was
element (IRE), a ZnS crystal with antireflective coating manufactured spe- selected as the measuring site, mainly because of the relatively thin stratum
cifically for this application. This ZnS crystal in contact with skin is insoluble corneum (15–30 μm) and the fact that this body part is little affected by
and non-toxic and does not present any risk for the user. The probe laser manual labor or exercise. Furthermore, this site was used because of its
deflection caused by the thermal lens is sensed by a proprietary custom-built relevance for the development of wearable versions of the device. The
detection system capable of resolving sub-μm movements of the probe laser, subjects wore a wristband reproducibly attachable to the D-Base for the

Communications Medicine | (2025)5:509 3

[Link] Article

duration of the session to improve data quality by standardizing the mea- study procedures or the D-Base and all rated as mild. All AEs were rated as
surement spot and pressure. non-serious and were resolved within the follow up period. No corrective
Before the start of the session, the wristband was fixed to the inner side actions were needed.
of the distal forearm, preferably close to the wrist of the participant, where it
remained for the duration of the session (Fig. 4). In some cases, the wrist- Signal processing
band was placed slightly further up the inner side of the distal forearm to The four machine-learning algorithms applied share the same pre-
improve skin contact with the measurement interface. In a few cases, the processing pipeline, i.e., the steps that transform the acquired raw data
wristband had to be moved slightly during the session to improve skin into features usable by the algorithms. Such pre-processing steps can be
contact with the measurement interface. In both scenarios, no statistically divided into two families: The first includes stateless steps that are applied
significant impact on the device performance was observed. independently to each individual measurement. The second is stateful and
Each measurement consisted of a reference measurement and a NIGM requires a calibration for each subject. Stateless methods include smoothing
measurement, in this order. To further analyze the variations of raw spectra in the wavenumber domain by a Savitzky-Golay filter18 and unit normal-
on identical measurement spots, the measurements were organized in sets of ization of measurements. This emphasizes relative amplitudes and spectral
four with a droplet of water as contact medium. The participant placed the patterns over absolute amplitudes of features. Processing steps are applied in
wrist on the device and did not lift it for four consecutive measurements. the following order: Smoothing, then unit normalization, followed by outlier
Each measurement took about 4 min, and measurements were performed in detection and a Net Analyte Signal (NAS) algorithm (see below).
succession, so a set of four lasted approx. 20 min. In between these sets, the Stateful methods are used for two purposes: Outlier detection and
spectrum of a drop of water was measured to monitor device stability. The discounting of non-glucose variations. A custom outlier detection algorithm
measurement time of approximately 4 minutes in this study was chosen for was developed, based on unsupervised embedding learning and subsequent
additional data collection and further analysis. Hotelling’s T-squared-statistics analysis of the spectra’s semantic
For each session, the participants arrived fasting at the study site in the representation19. It is calibrated for each subject based on the first session
morning. For the first hour, eight measurements were collected with fasting (for the second session) or on the first two sessions (for the third session) and
level glucose. Within the next four hours, up to 28 more measurements were used to support the identification and handling of outliers (“bad measure-
collected with changing glucose levels induced by a standardized breakfast at ments”) in subsequent sessions. Non-glucose variations are due to several
varying times within the session. factors, mainly caused by environment, instrument, and sample. They all
The reference values of the first session were known and used for vary across sessions, and it is impractical to model them individually. Our
calibration purposes. For the second and third session, the reference data approach is to use a NAS unsupervised learning technique that is trained for
points were blinded by the study site until all respective NIGM measure- each subject using the fasting level measurements available up to the session
ment results were provided for the session to the study site. Four proprietary to be predicted20. NAS transforms each measurement and presents the
machine-learning algorithms were tested for the determination of glucose machine learning algorithms with features that highlight glucose-specific
values, termed 1a, 1b, 2a and 2b. The main difference between the “1” and information.
“2” algorithms were their complexity and the available parameter space. The
difference between “a” and “b” algorithms was the recalibration with the first Algorithms
two measurement points of the test session and the glucose value span The four machine learning algorithms are all based on the Support Vector
during fasting glucose measurements for the “b” types. Consistently, these Regression (SVR) algorithm21,22. Compared to state-of-the-art neural net-
two values were not used to determine the accuracy of the “b” algorithms. works, it is well-suited for working with scarce data. We preferred the SVR
The results obtained by the NIGM algorithms were delivered once the algorithm to the de facto standard in spectroscopy, PLSR, because of its
respective session was completed. After the delivery of the NIGM mea- ability to focus on hard-to-model measurements by targeting them as
surement results was verified by the study site, the reference values were support vectors. In addition, the choice between different kernels during
transferred via email to DiaMonTech. This process was quality-assured with training allows further tuning for each subject. Model calibration for the
a four-eye principle by the study site to exclude any data leakage. second session is done on the first session, while the third session uses the
In general, for the non-invasive measurement no significant dis- first two sessions. Note that the first eight measurements of each session are
comfort or impact on tissue was reported by the subjects. This is to be not used to calibrate models, nor are they used to determine the accuracy:
expected as the tissue gets only warmed up below the user’s perception Rather, they are used for the unsupervised training of the NAS model.
threshold at the laser spot. The non-invasive device has passed the relevant The four algorithms were named 1a, 1b, 2a and 2b. “1” algorithms are
laser safety tests which are based on the MPE (Maximum Permissible limited to a linear SVR kernel because the physical phenomena under study
Exposure) for skin. Our device works below 50% of the maximum safety are linear in nature. “2” algorithms, on the other hand, extend the kernel
threshold. Overall, ten adverse events (AEs) were recorded. Five AEs hap- choice to better model the complexity of the real world. In addition, “b”
pened during the study participation without any relationship to the study algorithms, as opposed to “a” algorithms, employ a recalibration technique
or the device. The other five events were causally or possibly related to the that adds the first two points of the test session and the fasting glucose

Fig. 4 | Photographs showing the use of a wristband to ensure reproducible attachment to the D-Base. Left: open wristband; center: closed wristband; right: wristband
attached to D-Base.

Communications Medicine | (2025)5:509 4

[Link] Article

Fig. 5 | Time course of blood/tissue glucose for three individuals (R-010, R-024, measurements for each individual. The source data for Fig. 5 can be found in
R-032) at visits 2 and 3. Subject R-010 was a type-2 diabetes patient, subject R-024 a Supplementary Data
type-1 diabetes patient, and subject R-032 a healthy volunteer. Sample sizes were 26

change to the calibration. This technique exposes the algorithm to mea- Statistics and reproducibility
surements taken in the same environment, instrument and sample condi- Consensus error grid. The consensus error grid (CEG) is a risk-oriented
tions of the measurements to be determined, without revealing much of the graphical representation of the test results23. Glucose values are repre-
glucose test behavior, given its rather flat profile at the beginning. In total, up sented by the invasive reference value on the horizontal axis and the non-
to 38 (74) points were available for calibration purposes per user for the invasively measured value on the vertical axis. For an ideal measuring
second session (third session), while testing is computed on the up to 28 method, all data points would be found on the bold diagonal line. In
postprandial points per user. “b” algorithms, that use recalibration, have two reality, data points scatter along that line. The plot is divided by risk into
points less for testing that become two more points for calibration. Com- zones. Zones A and B are considered “safe” for the user. Any clinically
bining the results of all users, the total number of points available for testing accepted invasive test strip device needs to perform with 99% of all
of the most promising algorithm 2b was 896 (862) for the second session measured data in zones A and B, among other requirements (ISO
(third session). One user only performed one test session. All available test 15197:2013).
points are included in the statistical analysis and the consensus error grid.
Each predictive model is calibrated via cross-validation over the cor- Bland–Altman plot. A Bland–Altman plot is an established method to
responding training set. As a result, the best parameters across multiple assess the agreement between two measurement systems. It helps to
predictive models might vary. The features used by the predictive algorithms visualize and analyze effects of a fixed or proportional bias. It shows the
are solely the acquired absorbance spectra; no measurement metadata range within which ~95% of individual relative differences from the mean
(subject age, diabetes type, etc.) was used at this stage, as such information of test and reference system are found.
was disclosed to DiaMonTech only once the study was concluded.
The calibration is defined as “offline”, since algorithms to be tested on Mean Absolute Relative Difference. The Mean Absolute Relative
the second session are trained using data from the first session, while Difference (MARD) is one of the de facto standard performance
algorithms to be tested on the third session have the first and second session indicators of glucometers. It compares the measurement values of a
available as training set. This contrasts with “online” training, where an test device with those of a state-of-the-art reference device and is an
algorithm is trained incrementally as new data becomes available. important tool for the comparison of different methods, technologies
Over- and underfitting are relevant issues in the context of data scarcity and products in comparable test settings. For each value pair the
and complex environments like non-invasive glucose measurements. We Absolute Relative Difference (ARD) is calculated and then
have mitigated this risk using several methods: To prevent overfitting, we the arithmetic mean of the ARD of all value pairs is used to assess the
use calibration datasets that are representative of the predictive problem, as overall performance in a test session. We have thus decided to use
the model sees at least a full glucose tolerance test with the glucose fluctu- the MARD score for the estimation of the accuracy instead of other
ating. Our model choice, SVR, is less prone to overfitting22 compared to metrics like RMSE or MAD. This is common in the comparison of
other popular models (e.g., random forests, which have been tested in the different glucometer technologies and broadly used in specialized
field by other research groups) and requires less data to produce models journals.
which are not underfitted, compared to models such as neural networks.
Results
Reference device Contour Next Time course of blood/ISF glucose
As a reference device, the Contour Next with test strips and control solution Figure 5 exemplarily depicts for three participants the time course of BG
(Ascensia Diabetes Care Holdings AG, Basel, Switzerland) was used in this measured with the invasive reference device and with the non-invasive
study in combination with the lancing device Medisafe Finetouch II (Terumo technology D-Base using algorithm 2b. The shaded area around the refer-
Corporation, Tokyo, Japan), measuring the capillary BG. Although not an ence values represents an absolute relative difference of 20% compared to
official clinical reference method like venous BG analyzers, both are com- the state-of-the-art reference method. The D-Base measurements follow
mercially available devices with well characterized performance and are con- closely the general shape (absolute values and trend) of the reference curve
sidered representative for the standard measurement of a person with diabetes and are well within the acceptable band.
and sufficiently accurate as reference method in this study. We refrained from The best MARD values of 20.7% and 19.6% on the second and third
using a minimally invasive CGM sensor as a reference because of the lower sessions, respectively, were both obtained by algorithm 2b. The range of
precision and the potential delays with respect to blood glucose. BG mea- MARD values of the other algorithms was 22.9% to 35.3% (24.4% to 43.9%)
surements with Contour Next were performed in duplicate, and the average of for the second session (third session).
the duplicate was used as reference value for the NIGM device. As for the performance separated by diabetes type, Table 1 summarizes
No further medicinal products, human or animal tissues or their the MARD values (average, 95% confidence interval) for the best per-
derivatives, or other biologically active substances were used. forming algorithm 2b for the sessions 2 and 3.

Communications Medicine | (2025)5:509 5

[Link] Article

Table 1 | Performance of the non-invasive glucose measurement

Diabetes Type and Study Session Global Healthy Type 1 diabetes Type 2 diabetes
Session 2 MARD [CI 95%] 20.7% [16.7%, 24.5%] 15.8% [9.9%, 21.7%] 19.7% [13.4%, 26.0%] 23.2% [16.1%, 30.3%]
Session 3 MARD [CI 95%] 19.6% [16.6%, 22.9%] 12.7% [9.0%, 16.5%] 23.7% [16.3%, 31.1%] 19.4% [16.1%, 22.7%]

Fig. 6 | Consensus error diagrams of non-invasive glucose measurements obtained with algorithm 2b for the second (left) and third session (right). Sample size was 896
(session 2) and 863 (session 3). The source data for Fig. 6 can be found in Supplementary Data

Table 2 | Range of reference BG values in the test sessions, the four algorithms. The best performance was obtained for algorithm 2b,
globally and separated by diabetes type where Pearson correlation coefficients of 0.771 and 0.690 were achieved
for the second and third sessions, respectively.
Reference BG Global Healthy Type 1 Type 2
diabetes diabetes
The agreement rates between the reference device Contour Next and
the NIGM system for algorithm 2b showed the best results with 80.4%
Min 66 mg/dL 66 mg/dL 68 mg/dL 75 mg/dL
(second session) and 81.3% (third session) for the intervals of ±30% or
Max 368 mg/dL 202 mg/dL 368 mg/dL 323 mg/dL ±30 mg/dL for glucose concentrations ≥100 mg/dL or <100 mg/dL,
respectively.
A third method to analyze the measurement quality of the NIGM was
Clinical accuracy of the non-invasive method. The CEG allows a
the Bland–Altman plot, which indicates systematic influences and is
graphical representation of the test results, showing the comparison
depicted in Fig. 7. Again, algorithm 2b showed the best results out of the four
between invasive glucometer reference values and the NIGM measure-
algorithms. The 95% limit of agreement was ~±40%.
ments for all participants of the study. Figure 6 depicts the CEG of D-Base
vs. Contour Next for the second and third session of algorithm 2b. The
Discussion
data includes a wide range of reference BG values from 66 mg/dL to
For clinical validation of new measurement technologies, measurement
368 mg/dL, as summarized in Table 2.
results are compared with reference values from a state-of-the-art gluc-
For algorithm 2b, 97.5% (98.6%) of the values are within clinically
ometer during times of changing glucose levels. To classify clinical valida-
accurate zones A + B in the second session (third session). Only one
tions of glucose measuring techniques, it is helpful to distinguish between
measurement fell into zone D, none in zone E. This performance is close to
two different approaches.
one of the requirements for invasive blood glucometers according to ISO
In this manuscript, we refer to approaches as “retrospective” when the
15197:2013.
reference values are available before the measurement results are provided.
For the other three algorithms, values in A + B of the CEG range from
These studies usually divide their datasets into a calibration and a validation
89.1% (83.8%) to 96.5% (97.9%) for the second session (third session), with
dataset with varying segmentation of points. It is a common approach in
0% (0%) up to 1.1% (1.2%) in zone D and none in zone E. The performance
early feasibility and explorative studies.
of all algorithms indicates the base measurement function of the technology,
In contrast, an approach is defined as “prospective” when the non-
independent of the algorithm choice.
invasive measurement results are provided before knowing the reference
values (following an initial calibration). This study type is common for more
Analytical accuracy of the non-invasive method. The Pearson cor-
mature technologies, simulating real-life scenarios.
relation coefficient is an indicator for the measurement capabilities of a
As the retrospective data scientific problem to be solved is simpler, such
glucose monitor and a value > 0.5 was designated to indicate an ability to
approaches tend to yield a better Mean Absolute Relative Difference
follow changes of BG concentrations. This boundary was set as a success
(MARD) than prospective ones.
criterion for this study and could be shown on both visits for three out of

Communications Medicine | (2025)5:509 6

[Link] Article

Fig. 7 | Bland–Altman plot for all measurements of the second (left) and third (right) session as obtained with algorithm 2b. Sample size was 896 (session 2) and 863
(session 3). The source data for Fig. 7 can be found in Supplementary Data

In a previous study (“Study 100”) with 100 volunteers, we demon- many advantages compared to (minimally) invasive methods, the benefits
strated the feasibility of the technology. A MARD of 12% was achieved. That for a device based on the demonstrated accuracy are sufficient for
study has been done in a retrospective manner and indicated the feasibility adjunctive use.
of an NIGM approach based on MID-IR technology24. The measurements were based on the glucose absorption in the mid-
With the improved test device, the study we report here was conducted infrared spectral region, at wavelengths between ~8 and 11 μm. This “glu-
in a prospective fashion. For all utilized metrics, algorithms 2 showed better cose fingerprint“ (Fig. 1) is highly specific for the glucose molecule. Since the
results than the corresponding algorithm 1, and algorithms b outperformed spectral features of glucose in this spectral range arise from fundamental -C-
the respective algorithms a. Algorithm 2b delivered the best performance, O- and -O-H vibrations, their characteristics are more reliable than the
closely followed by algorithm 2a. This is especially impressive, as the latter overtones (harmonics) that are found in the near infrared between ~1 and
uses no recalibration from the session day. 3 μm, at the expense of more sophisticated light sources, optics and detec-
The performance of algorithm 2b for the second and third sessions in tion technology.
MARD is comparable to early-stage CGM from Medtronic and Dexcom The external cavity-tuned QCLs used in the D-Base were originally
that the FDA approved for adjunctive use in 2005 and 2006, respectively25,26. developed as powerful and broadband tunable MID-IR sources for a wide
To our knowledge, no truly non-invasive glucometer has reached this range of scientific applications. A QCL optimized for the requirements of
MARD in a similar prospective clinical study setting. BG measurements could improve the performance as well as reduce the size
In their early days, minimally invasive CGM had the disruptive and costs. Broad tunability over the entire glucose fingerprint (Fig. 1) is not
advantage of augmenting, not replacing, state-of-the-art finger-prick mea- an essential requirement but can be substituted by using selected wave-
surements with a more frequent monitoring of blood sugar values. A truly lengths in the fingerprint spectral range. Using optimized arrays of QCL
non-invasive system promises a similar disruptive nature to the state-of-the- emitters that can be controlled individually with sufficient emission power
art methods of BG measurements with its ability of pain free and harmless could significantly reduce the measurement time to well below a minute and
use without consumables. Under the premise of a similar use as an enable further miniaturization. Based on the gathered knowledge, Dia-
adjunctive device without direct therapeutic consequence, it is hence rea- MonTech developed the first miniaturized dedicated glucose-QCL-array in
sonable to compare the performance to those early-stage CGM systems. cooperation with a German laser manufacturer. These arrays, which open
The performance of CGMs has improved over the years, so that current the possibility for integration into wearable technology, could mitigate the
generations of commercially available CGM systems are cleared for non- drawbacks of the off-the-shelf QCLs used in the D-Base.
adjunctive use. In a recent comparison of different invasive self-monitoring Based on these QCL arrays and learnings from the D-Base develop-
glucometers, MARD values from about 2.3% to about 20% were reported27. ment, DiaMonTech has already miniaturized optics and electronics towards
For modern CGM systems, MARD values of up to 20% have been reported a handheld version “D-Pocket” as visualized in Fig. 8. The detection method
in an independent study28, yet MARD values of around 10% have been in the D-Pocket is essentially the same as for D-Base. The D-Pocket has the
published repeatedly in manufacturer-funded studies. As there are no footprint of a common smartphone. It is currently a prototype under
relevant NIGM products established on the market, a similar scientific development. The D-Pocket is planned as an individual companion for
comparative evaluation of the accuracy of truly non-invasive glucometers is people with diabetes with the capacity for 30–50 measurements with one
not yet available. Thus, such a comparison can only be made based on data battery charge.
provided by research groups and manufacturers. A small number of data points exhibit substantial deviations between
The feasibility of NIGM approaches is supported by a clinical test of the non-invasive measurement from the reference and are part of the
another non-invasive system based on Raman spectroscopy reported investigations in the ongoing development work. Lab testing with the
recently29. This NIGM system reached a MARD value of 19.2%. This D-Base as well as further analysis of data from this study suggest that sources
compares well with the accuracy obtained with the system tested in this for signal instabilities for this non-invasive glucose measurement technique
study, yet in a different study setup with a substantially longer calibration are the device itself, skin inhomogeneity, movement artefacts and artefacts
period of 26 days and home use without dedicated glucose swings. from skin-care products. We expect to cope with the inhomogeneity of skin
and influence of skin-care products by a more detailed analysis of the raw
Conclusions infrared spectra through modulation of the MIR pump laser. The aim is to
We report here a clinical study on the evaluation of a truly non-invasive identify sweat glands, scars or hair, substances on the surface of the skin and
glucose sensor reaching an accuracy of MARD < 20% which is comparable other potential sources for measurement inaccuracies. Reduction of the
to FDA-approved early stage CGMs. As a non-invasive measurement holds

Communications Medicine | (2025)5:509 7

[Link] Article

we estimate based on lab tests a lifetime of >3 years limited only by standard
wear and tear of an optical measurement device and a battery.
For the handheld version, preliminary cost estimates can be made
based on a scalable production of the main component, the quantum cas-
cade laser, the associated optics, and a lifetime of the device of several years.
As an outcome, the costs over time for non-invasive glucose measurement
using this technology will be strongly competitive to the costs for minimally
invasive sensors.

Data availability
The complete source data for Figs. 5, 6 and 7 can be found in one Excel sheet
in Supplementary Data. The complete report of the clinical test at the IfDT
can be requested from the corresponding author.

Received: 25 October 2024; Accepted: 3 November 2025;

References
1. International Diabetes Federation. IDF Diabetes Atlas: 11th Edition
[Link] (International Diabetes Federation, 2025).
2. Petry, S. F., Petry, F. W., Petry, J. K., Gäth, S. & Heinemann, L.
Diabetes technology and waste: a real-world study in a specialized
practice in Germany. J. Diab. Sci. Technol. 2024, 1–9 (2024).
3. Gonzales, W. V., Mobashsher, A. T. & Abbosh, A. The progress of
glucose monitoring—a review of invasive to minimally and non-
invasive techniques, devices and sensors. Sensors 19, 800 (2019).
4. Pleus, S. et al. Documentation of skin-related issues associated with
CGM use in the scientific literature. Diabetes Technol. Ther. 21.
Fig. 8 | Visualization of the handheld glucometer “D-Pocket”. With a form factor [Link] (2019).
of a smartphone it will be portable and suitable for everyday life usage. 5. Di Fillippo, D., Sunstrum, F. N., Umar Khan, J. & Welsh, A. W. Non-
invasive glucose sensing technologies and products: a
comprehensive review for researchers and clinicians. Sensors 23,
measuring time from a few min to the order of 10–20 s greatly reduces the 9130 (2023).
risk of movement artefacts. 6. Hina, A. & Saadeh, W. Noninvasive blood glucose monitoring systems
The data collected in this study will be used to further improve the using near-infrared technology—a review. Sensors 22, 4855 (2022).
algorithms with the aim of simplifying the calibration procedure and 7. Scholtes-Timmerman, M. J., Bijlsma, S., Fokkert, M. J., Slingerland, R.
improving performance. As demonstrated, algorithm 2a, without recali- & van Veen, S. J. F. Raman spectroscopy as a promising tool for
bration on the session day, performed only slightly worse than the best noninvasive point-of-care glucose monitoring. J. Diab. Sci. Technol.
algorithm, giving reason to believe that the algorithms can further be refined 8, 974–979 (2014).
by reducing calibration requirements. With more data available, algorithms 8. Pandey, R. et al. Noninvasive monitoring of blood glucose with Raman
could be enhanced for a global calibration model for little to no (re-)cali- spectroscopy. Acc. Chem. Res. 2017 50, 264–272 (2017).
bration. We expect to collect this data with the handheld version “D-Pocket” 9. Lundsgaard-Nielsen S. M. et al. Critical-depth Raman spectroscopy
that is currently finalized. The “D-Pocket” is planned as an individual enables home-use non-invasive glucose monitoring. PLoS One 13.
companion for people with diabetes. As such, a trial with a larger number of [Link] (2018).
devices used individually by the patients is planned. For the future, a 10. Pleitez, M. A., von Lilienfeld-Toal, H. & Mäntele, W. Infrared
wearable form factor based on a further miniaturized technology is in the spectroscopic analysis of human interstitial fluid in vitro and in vivo
concept phase. using FT-IR spectroscopy and pulsed quantum cascade lasers (QCL):
In summary, we have successfully demonstrated the accuracy of a non- establishing a new approach to non-invasive glucose measurement.
invasive glucose measurement technology. The study procedure was Spectrochim. Acta Part A 85, 61–65 (2012).
designed in a prospective fashion, i.e., measurements were done without 11. Liakat, S. et al. Noninvasive in vivo glucose sensing on human
prior knowledge of the reference values, aiming to a realistic situation of a subjects using mid-infrared light. Biomed. Opt. Express 5, 2397–2404
diabetes patient. Based on MARD values, the Pearson correlation coeffi- (2014).
cient, the Bland and Altman analysis, and the consensus error grid analysis, 12. Kasahara, R., Kino, S., Soyama, S. & Matsuura, Y. Noninvasive
this NIGM system exhibits the ability to follow changes of BG concentra- glucose monitoring using mid-infrared absorption spectroscopy
tions. The demonstrated performance opens new perspectives for better based on a few wavenumbers. Biomed. Opt. Express 9, 289–302
management of diabetes through realization of a reliable and accurate (2017).
product for everyday use in the near future. 13. Uluç, N. et al. Non-invasive measurements of blood glucose levels by
The device developed and used in this clinical validation (D-Base), and time-gating mid-infrared optoacoustic signals. Nat. Metab. 6,
the handheld version under development (D-Pocket), do not require con- 678–686 (2024).
sumables such as disposable sensors, except for the test strips used for the 14. Basu, A. et al. Time lag of glucose from intravascular to interstitial
calibration procedure with an invasive glucometer as described above. Yet, compartment in humans. Diab. 62, 4083–4087 (2013).
both allow an unlimited number of measurements to improve the glycemic 15. Bauer A., et al. IR-spectroscopy of skin in vivo: optimal skin sites and
condition of the diabetes patient. The issue of lifetime of these optical properties for non-invasive glucose measurement by photoacoustic
sensors (e.g., as compared to the maximum lifetime of CGM sensors of and photothermal spectroscopy. J. Biophotonics11. [Link]
about 14 days) is not relevant. As for the lifetime of the entire sensor device, 10.1002/jbio.201600261 (2018).

Communications Medicine | (2025)5:509 8

[Link] Article

16. Pleitez, M. A. et al. Photothermal deflectometry enhanced by total maintaining the D-Bases and making the study possible. We would also like
internal reflection enables noninvasive glucose monitoring in human to thank the IfDT for conducting this study and for the pleasant collaboration.
epidermis. Analyst 140, 483–488 (2015).
17. Hertzberg, O., Bauer, A., Küderle, A., Pleitez, M. A. & Mäntele, W. Author contributions
Depth-selective photothermal IR spectroscopy of skin: potential M.K., S.J., P.L. and M.S. prepared the non-invasive measurement tech-
application for non-invasive glucose measurement. Analyst 142, nology for clinical validation and the general study design. L.C. and V.L. were
495–502 (2017). responsible for the development of the signal processing and machine
18. Savitzky, A. & Golay, M. J. Smoothing and differentiation of data by learning algorithms. T.L. supervised the study. W.M. wrote the first draft of
simplified least squares procedures. Anal. Chem. 36, 1627–1639 the manuscript. All authors edited the manuscript.
(1964).
19. Mujica, L. E., Rodellar, J., Fernandez, A. & Güemes, A. Q-statistic and Competing interests
T2-statistic PCA-based measures for damage assessment in M.K. is CTO, S.J. is COO, T.L. is CEO, and W.M. is CSO of DiaMonTech AG.
structures. Struct. Health Monit. 10, 539–553 (2010). T.L. and W.M. are co-founders of DiaMonTech AG. L.C., P.L., V.L. and M.S.
20. Lorber, A., Faber, K. & Kowalski, B. R. Net analyte signal calculation in are employees of DiaMonTech.
multivariate calibration. Anal. Chem. 69, 1620–1626 (1997).
21. Awad, M. & Khanna, R. Support vector regression. In: Efficient Additional information
learning machines. Apress. [Link] Supplementary information The online version contains
5990-9_4 (2015). supplementary material available at
22. Bishop C. M. & Nasrabadi N. M. Pattern Recognition and Machine [Link]
Learning (Springer, 2006).
23. Parkes, J. L., Slatin, S. L., Pardo, S. & Ginsberg, B. H. A new Correspondence and requests for materials should be addressed to
consensus error grid to evaluate the clinical significance of Werner Mäntele.
inaccuracies in the measurement of blood glucose”. Diab. Care. 23,
1143–1148 (2000). Peer review information Communications Medicine thanks Anas El Fathi,
24. Lubinski, T., Plotka, B., Janik, S., Canini, L. & Mäntele, W. Evaluation of Pradyut Kumar Sanki and the other, anonymous, reviewer(s) for their
a novel noninvasive blood glucose monitor based on mid-infrared contribution to the peer review of this work. A peer review file is available.
quantum cascade laser technology and photothermal detection. J.
Diab. Sci. Technol. 15, 6–10 (2021). Reprints and permissions information is available at
25. Food and Drug Administration. Summary Of Safety And Effectiveness [Link]
Data - Medtronic Guardian RT (PMA P980022/S011) Available at:
[Link] Publisher’s note Springer Nature remains neutral with regard to
pdf; Accessed date: 29 May 2024. jurisdictional claims in published maps and institutional affiliations.
26. Food and Drug Administration. Summary Of Safety And Effectiveness
Data - DexCom#STS# Continuous Glucose Monitoring System (PMA Open Access This article is licensed under a Creative Commons
P050012) Available at: [Link] Attribution-NonCommercial-NoDerivatives 4.0 International License,
pdf5/[Link]; Accessed date: 29 May 2024. which permits any non-commercial use, sharing, distribution and
27. Freckmann, G. et al. Mean absolute relative difference of blood reproduction in any medium or format, as long as you give appropriate
glucose monitoring systems and relationship to ISO 15197. J. Diab. credit to the original author(s) and the source, provide a link to the Creative
Sci. Technol. 16, 1089–1095 (2022). Commons licence, and indicate if you modified the licensed material. You
28. Heinemann, L. et al. Benefits and limitations of MARD as a do not have permission under this licence to share adapted material
performance parameter for continuous glucose monitoring in the derived from this article or parts of it. The images or other third party
interstitial space. J. Diab. Sci. Technol. 14, 135–150 (2020). material in this article are included in the article’s Creative Commons
29. Pors, A. et al. Accurate post-calibration predictions for noninvasive licence, unless indicated otherwise in a credit line to the material. If material
glucose measurements in people using confocal Raman is not included in the article’s Creative Commons licence and your intended
spectroscopy. ACS Sens. 8, 1272–1279 (2023). use is not permitted by statutory regulation or exceeds the permitted use,
you will need to obtain permission directly from the copyright holder. To
Acknowledgements view a copy of this licence, visit [Link]
We thank Raja Bilal Ahmed, Philippa Clark, Jenny Kegel, Johannes Lutz, nc-nd/4.0/.
Yigit Mahsereci, Nikita Malyshev, Lars Melchior, Alexander Mittelstädt,
Daniel Rojas (all at DiaMonTech) as well as /dev for preparing and © The Author(s) 2025

Communications Medicine | (2025)5:509 9