Non-Invasive Optical Glucose Sensing Physics

Non-Invasive Optical Glucose Sensing Physics

Reading glucose with light through intact skin is hard for a physical reason. Glucose makes up only about 0.07 to 0.1% of blood plasma, while water makes up about 90% [1], and water, pigments and proteins in tissue all absorb light at the wavelengths a sensor would use [2]. The glucose signal is tiny, and the background it sits on keeps moving. As of October 2026, the FDA says it has not authorized any smartwatch or smart ring that measures or estimates blood glucose on its own [3]. For a tour of the individual methods, see our article on optical and spectroscopic glucose sensing.

What happens to light inside skin

Light that enters skin is either absorbed or scattered. Absorption removes a photon and hands its energy to a molecule. Scattering only changes the photon's direction. Total attenuation, how quickly a beam fades, is the sum of an absorption coefficient and a reduced scattering coefficient [1]. The reduced version adjusts for the fact that tissue scatters light mostly forward [4].

In skin, the main absorbers are blood, water, melanin, fat and yellow pigments, while structures large and small inside cells and tissue do the scattering [5].

On paper, absorption follows the Beer-Lambert law: absorbance rises with the concentration of a substance, with how strongly it absorbs, and with the distance light travels through the sample, called the path length [2]. That formula ignores scattered light [2]. In skin, scattering is the rule.

Path length: the number nobody controls

In a laboratory cell, path length is fixed. Heise and colleagues measured plasma samples through a constant 1 mm path [4]. In skin, photons take a random walk, and how deep they reach depends on the wavelength and on the tissue's absorption and scattering coefficients [4]. To read glucose from lip spectra, the same group had to scale glucose's laboratory spectrum by a simulated penetration depth at every wavelength [4].

Wavelength choice trades depth against chemistry. Between roughly 600 and 1,300 nanometers, tissue absorbs little compared with how much it scatters, so light can pass through a fingertip [4]. At longer near-infrared (NIR) wavelengths, where glucose's more useful bands sit, so little light gets through that devices collect diffuse reflection instead [4].

Skin also varies between people. A 2023 Raman study cites a dead outer layer (stratum corneum) 166 ± 40 micrometers thick at the base of the thumb, so its probe, focused at a fixed 280 micrometers, collected a different mix of dead and living tissue from each person, which changes the raw signal-to-noise ratio [6]. Contact matters too. Probes pressed on soft tissue often leave a pressure mark, which led one team to build a non-contact probe that samples a larger volume [7]. In an animal study, just repositioning the probe produced significant scatter in glucose predictions [4].

Why water drowns the glucose signal

In the NIR range, water has two strong absorption peaks, about 1,350 to 1,520 nm and 1,790 to 2,000 nm, so glucose work uses the windows between them [2]. Glucose shows up there only through overtone and combination bands, weaker echoes of its main vibrations [4]. Those bands are broad and overlap with what other tissue components absorb [7]. Heise's review lists the result plainly: a tiny glucose absorbance on a high, variable water background, with baseline shifts, instrument drift, low sensitivity and poor precision [4].

Here's what surprised us: part of what changes when glucose rises is the water itself. Jensen and colleagues measured how adding glucose changes water's own absorption, alongside glucose's spectrum [8]. Heise's group went further and built "water displacement" into the noise term of its calibration model [4].

Temperature, fat and a moving background

Water's spectrum also shifts with temperature. Jensen's team measured pure water and a 1 g/dL glucose solution from 30 to 42 °C in 2-degree steps [8]. The temperature effect was strongest near water's big absorption bands, with several isosbestic points, wavelengths that don't change with temperature, and flatter stretches between them [8].

Holding the skin at a set temperature doesn't solve it. With the lip contact area kept at 37 °C, Heise's group still found temperature gradients; the largest sources of spectral variation were water and its temperature dependence, plus signatures of fat under the skin [4]. An MIT team writing in 2020 notes that in living tissue, changes in temperature and contact pressure easily swamp weak NIR glucose signals [7]. Some designs add skin and room temperature sensors and feed both readings into the glucose model [1].

Scattering as signal, not just noise

Scattering can carry glucose information too. Glucose changes the refractive index of the fluid between cells, and the mismatch between that fluid and the cells helps set the reduced scattering coefficient [9]. In 1994, Maier and colleagues built a frequency-domain NIR spectrometer precise enough to detect scattering changes across the physiological glucose range [9].

In a 2000 study of 18 people given a 75 g oral glucose load, scattering changes followed glucose acceptably in 8 of 10 measurements in healthy volunteers and 19 of 26 in people with type 2 diabetes, and the authors couldn't yet explain why it failed in the rest [10]. Arnold and Small call this an indirect approach, since it measures glucose's effect on something else rather than the molecule itself, and say selectivity limits indirect and direct methods alike [11].

Calibration and signal-to-noise

Because no single wavelength isolates glucose, optical systems rely on multivariate calibration: statistical models fitted to many spectra paired with reference blood glucose values [4]. With a signal this weak, a model can latch onto anything that happens to track glucose during a study. The MIT team, with co-authors from Samsung Advanced Institute of Technology, pointed out that earlier in-vivo Raman reports showed statistical correlation but never glucose's actual Raman peaks, leaving chance correlation open [7]. In three glucose clamp experiments in pigs, they observed a glucose peak at 1,125 cm⁻¹ whose intensity rose in step with glucose [7].

In a 2023 home study of 160 people with diabetes, which its authors called the largest of its kind, a portable Raman device from RSP Systems held its accuracy for 15 days after calibration; root-mean-square error rose only from 1.68 to 1.84 mmol/L [6]. Overall mean absolute relative difference (MARD, the average percentage gap from reference readings) was 19.1%, and 58.8% for reference values below 3.9 mmol/L [6]. Each person's model needed about 20 latent variables (hidden patterns combined into one estimate), which the authors read as a sign of a small glucose signal inside a widely varying spectrum [6]. Six authors worked for RSP Systems, and the device takes intermittent readings without the trend correction CGMs use [6].

What the record shows

A 2022 review found 28 noninvasive optical glucose products, none approved by the FDA, and called accuracy over time the major analytical challenge for all of them [12]. It also flagged skin pigmentation, which the FDA has warned affects pulse oximeter accuracy, as a possible problem for optical glucose monitors [12]. The 2023 Raman study saw similar accuracy across skin colors, but its darkest and lightest groups had only 2 and 3 people [6]. Our article on non-invasive glucose sensing technologies covers the non-optical methods.

Choosing an FDA-authorized meter or CGM is a decision to make with your care team. The physics question still open is whether any optical design can keep its calibration across skin types, temperatures and months of wear.