Why non-invasive glucose monitoring is hard | Empirical Health

New: 100 biomarkers for $190 server-island-start<br>Why non-invasive glucose monitoring is hard<br>Brandon Ballinger · Jul 20, 2026

Continuous glucose monitoring on Apple Watch and other smartwatches has been “5-7 years away” for roughly a decade.

In that time, we’ve seen the launches of many other health features, including ECG (2018), blood oxygen (2020), atrial fibrillation history (2022), sleep apnea (2024), and hypertension (2025). Many of these are even FDA-cleared:

Wearable manufacturers haven’t given up. Apple Watch’s glucose project was first reported in 2017, shrunk a tabletop-sized prototype to the size of an iPhone in 2023, and got a new leader in 2026 whose background suggest commercialization is approaching.

I trained one of the first deep neural networks to detect signs of diabetes with consumer health sensors. This post will explain what makes glucose sensing so hard, what hardware and machine learning techniques have been tried, and try to describe which research techniques are actually feasible on a consumer device like an Apple Watch, Pixel, Oura, or Samsung Watch.

Let’s start by explaining why an already-launched feature, blood oxygen sensing, actually works in practice using relatively cheap optical sensors.

A less hard example: blood oxygen sensing

Blood oxygen saturation launched with the Apple Watch and other wearables in 2020. The basic technology requires shining two wavelengths of light through tissue, typically 660 nm (red) and 940 nm (infrared).

This works is for three fundamental reasons:

Hemoglobin is abundant (about 14 g/dL in adult blood).

Oxygenated and deoxygenated hemoglobin have distinct spectra. At 660 nm, deoxyhemoglobin absorbs about ten times more strongly than oxyhemoglobin. At 940 nm, the relationship flips, which means the absorbsion ratio gives you oxygen saturation nearly directly.

The PPG signal has both AC (pulsing arterial blood) and DC (skin, bone, venous blood) components. Subtracting DC from AC isolates what’s happening in blood.

The result is that you can measure blood oxygen on the wrist with relatively inexpensive hardware: “just” two cheap LEDs and a photodetector.

Why glucose is inherently harder than hemoglobin

On the left, light from an LED enters the skin, follows a banana-shaped path down through the epidermis, dermis, and blood vessels (which pulse, enabling the AC/DC trick), and scatters back up to a photodetector a few millimeters away, so it crosses every layer twice. On the right, the absorption spectra of the four molecules the light interacts with, over 600 to 1900 nm: oxygenated and deoxygenated blood look different at 660 and 940 nm, water dominates tissue absorption with a huge band near 1450 nm, and glucose is spread through every layer but is about 1000x too weak to see without magnification.

Among all three dimensions above, glucose is a harder molecule: glucose is 1400x less concentrated than hemoglobin, has no isolated light wavelengths, and the AC/DC trick doesn’t work:

PropertyGlucoseHemoglobin (Oxygen)Concentration in blood ~90 mg/dL~14,000 mg/dLSpectra No isolated absorption peaks in visible/NIR; weak C-H overtones masked by waterDistinct oxygenated/deoxygenated peaks at 660 and 940 nmAC/DC Trick Not effective since glucose is everywhere and not pulsedHighly effective<br>Glucose doesn’t have isolated absorption peaks in the visible or near-infrared. It has C-H bond vibrations around 1500-1850 nm and 2050-2330 nm, but those are also absorbed strongly by water in those same regions. The signal from water is about 1000x stronger than glucose, leading to an awful lot of noise. Furthermore, glucose isn’t confined to pulsing blood since it’s found in plasma, red cells, interstitial fluid, and so on, which means the AC/DC trick doesn’t work. This is why despite 30 years of research, there’s still no FDA-cleared non-invasive glucose monitoring technology.

Non-invasive glucose monitoring techniques people have tried

Medical research has tried many approaches noninvasive glucose sensing, which cluster into six categories: NIR, MIR, Raman, photoacoustic, fluorescence, and PPG + machine learning.

The main accuracy metric used for glucose sensing is mARD: mean Absolute Relative Difference . To calculate mARD, you take multiple readings from a CGM and compare it to a reference. Each reading has a relative difference of (reading-reference)/reference; if you take the mean of the absolute value of these differences, that’s mARD. A mARD of 10% or less is considered accurate. Modern CGMs (e.g., FreeStyle Libre 3, Dexcom G7) achieve an mARD of about 8%.

Near-infrared spectroscopy

NIR (700 to 2500 nm) is the obvious starting point since it’s same basic tech as oxygen and heart rate sensing. We covered the problems with concentration, specta, and AC/DC above. While small proof-of-concept studies of NIR techniques report 5–8% mARD, broader prospective human studies cluster...