Can Your Apple Watch Predict Your Blood Sugar?

A personal research project exploring non-invasive glucose monitoring

David Bolis  •  June 2026 E:davidbolis@tapaway.com.au

Over 500 million people worldwide live with diabetes. Many of them wear a continuous glucose monitor (CGM) — a small sensor inserted under the skin that reads blood glucose every five minutes. It is life-saving technology. It is also a needle, a wearable patch, a recurring prescription, and a constant physical reminder that the body requires external monitoring to stay safe.

I started wondering: what if they didn’t need it?

Not because the technology is bad — it’s remarkable. But because the Apple Watch that many diabetics already wear on their wrist is quietly measuring heart rate, heart rate variability, skin temperature, oxygen saturation, respiration rate, and movement — every single minute of every day. That’s a lot of signal. And glucose, it turns out, leaves fingerprints in all of those signals.

This is the story of a research project I’ve been building to find out how close we can get.

The Science Behind the Idea

Blood glucose does not exist in isolation inside your body. When your blood sugar rises or falls, it triggers a cascade of physiological responses — changes in how your heart beats, how your autonomic nervous system responds, how your peripheral blood vessels dilate, even how warm your skin feels. These are not random side effects. They are measurable.

Research groups have been exploring this space for years. Studies have shown correlations between:

•       Heart rate variability (HRV) and glucose fluctuations — particularly hypoglycemic events

•       Resting heart rate elevation during hyperglycemia

•       Skin temperature changes linked to autonomic nervous system responses to glucose shifts

•       Respiratory patterns affected by metabolic state

•       Movement and activity patterns that both influence and are influenced by glucose levels

None of these correlations are perfect. A single heart rate reading tells you almost nothing about glucose. But when you stack multiple signals together over time — looking at patterns, trends, and interactions — a picture begins to emerge. That’s where machine learning becomes the right tool.

The hypothesis is straightforward: a model trained on enough people with known glucose values, using the same sensor suite that’s already on millions of wrists, should be able to learn those patterns well enough to make useful predictions.

Why Now? Why Apple Watch?

This idea is not new. Academic research into non-invasive glucose monitoring goes back decades. So why hasn’t it been solved?

Part of the answer is that most prior research tried to measure glucose directly through the skin using near-infrared light or other optical methods. This is extremely hard — the glucose signal is tiny, swamped by noise from water, hemoglobin, and other molecules, and highly variable between individuals. Companies have spent hundreds of millions of dollars on this approach with limited success.

The approach I’m exploring is different. Instead of trying to measure glucose directly, the goal is to infer it indirectly — using the physiological signals that glucose changes produce. This is harder to generalize, but much easier to personalize. And personalization, it turns out, is the key.

The Apple Watch is a uniquely interesting platform for this research for three reasons:

•       It is already on the wrist of hundreds of millions of people, many of whom are diabetic or pre-diabetic.

•       It collects a richer set of physiological signals than any previous consumer wearable — and with each hardware generation, that set expands.

•       It stores years of continuous, timestamped health data in Apple’s HealthKit, which means research doesn’t have to start from zero.

Ten years of Apple Watch data is sitting in a database on my Mac right now. Over a million data points. That’s not nothing.

The Project: What I’m Building

The project has two phases, and I’m currently between them.

Phase 1: A General Model

The first phase uses a clinical research dataset — one that pairs continuous glucose monitor (CGM) readings with wrist-worn sensor data from diabetic patients over multiple weeks. The goal here is to build a baseline model: does the relationship between wrist sensor signals and blood glucose generalize across people? How accurate can a general model get?

The answer from the literature is: somewhat. General models trained on diverse populations tend to achieve mean absolute relative differences (MARD) in the 15–25% range — not good enough to replace a CGM, but good enough to be useful as a secondary signal or trend indicator.

But “somewhat” is not the goal. The goal is: good enough for diabetics to not need a needle in their arm.

Phase 2: A Personal Model

This is where it gets interesting. Phase 2 is built around a volunteer participant who uses a Dexcom G7 CGM — meaning their Apple Watch health export includes actual blood glucose readings synced directly to Apple Health alongside the full wrist sensor history. That pairing is a ground-truth training set specific to one person’s physiology.

The hypothesis for Phase 2 is that a model fine-tuned on an individual’s data specifically — learning their personal relationships between HRV and glucose, their skin temperature baseline, their post-meal heart rate patterns — will dramatically outperform a general model. The physiology is individual. The model should be too.

This is the part of the project I’m most excited about and most uncertain about. The uncertainty is honest — this might not work.

The iOS App

Research that lives only on a MacBook is not useful to anyone wearing an Apple Watch. So alongside the machine learning pipeline, I’ve been building an iOS app — the thing that would actually put predictions on a person’s wrist.

The app reads live sensor data directly from HealthKit, displays real-time glucose predictions, allows users to log corrections when the prediction is wrong (which feeds back into model improvement), and supports both mg/dL and mmol/L for users anywhere in the world.

The correction logging feature is particularly important. The model will be wrong sometimes — especially early on. Every correction is a labeled data point that makes the next version of the model more accurate. The app is designed to get smarter the more someone uses it.

Honest Limitations

I want to be clear about what this project is and is not.

This is not a medical device. Nothing produced by this research should be used to make treatment decisions without consulting a healthcare professional. The goal is to explore whether the signal exists and how strong it is — not to replace clinical tools prematurely.

There are also real limitations to the indirect inference approach. It will likely never be as accurate as a CGM for acute events — a rapid hypoglycemic drop during exercise, for example, may not produce enough physiological signal fast enough for a wrist sensor to catch in real time. A CGM measures glucose directly; this approach infers it. That gap matters.

But “good enough to reduce Dexcom reliance significantly” is a much lower bar than “good enough to replace it entirely,” and it’s a bar that seems worth trying to clear.

Where This Goes

The immediate next step is getting the clinical training dataset and running the Phase 1 model. That will tell me whether the baseline signal is real and how well it generalizes.

After that, Phase 2 begins — fine-tuning on an individual’s personal data and watching the error rate drop. If it drops far enough, the app becomes something a diabetic person actually uses day to day alongside — and eventually instead of — their CGM.

Longer term, if the personalized model works well, the logical next step is better hardware — a wrist device with raw optical sensors that can capture the photoplethysmography waveforms that research suggests carry the strongest non-invasive glucose signal. That’s a hardware project for a later chapter.

For now, the question is simple: is the signal there? I think it is. I’m going to find out.

Glossary of Terms

A quick reference for the medical and technical terms used throughout this article.

Term	Definition
CGM	Continuous Glucose Monitor. A wearable medical device with a small sensor inserted just under the skin that measures blood glucose levels every few minutes and transmits readings wirelessly.
HRV	Heart Rate Variability. The variation in time between consecutive heartbeats. A key indicator of autonomic nervous system activity and metabolic state, and one of the signals most correlated with glucose fluctuations.
MARD	Mean Absolute Relative Difference. The standard accuracy metric for glucose monitoring. It measures the average percentage error between predicted and actual glucose values. Clinical CGMs typically achieve MARD below 10%.
ML	Machine Learning. A branch of artificial intelligence in which a model learns patterns from data rather than being explicitly programmed with rules. Used here to find the relationship between wrist sensor signals and blood glucose.
mg/dL	Milligrams per deciliter. The standard unit for blood glucose concentration used in the United States. Normal fasting range: 70–99 mg/dL.
mmol/L	Millimoles per litre. The standard unit for blood glucose used outside the US (UK, Australia, Canada, etc.). Normal fasting range: 3.9–5.5 mmol/L. To convert from mg/dL, divide by 18.
SpO2	Peripheral oxygen saturation. A measure of how much haemoglobin in the blood is carrying oxygen, expressed as a percentage. Measured non-invasively by the Apple Watch using light sensors on the wrist.
PPG	Photoplethysmography. An optical technique that detects blood volume changes in tissue using light. The raw PPG waveform from a wrist sensor carries subtle signals that research suggests are correlated with glucose levels.
HealthKit	Apple’s centralised health data platform on iPhone and Apple Watch. It stores all sensor readings — heart rate, HRV, steps, sleep, and more — in a secure, timestamped database on the device.
Hyperglycemia	High blood glucose. Generally defined as above 180 mg/dL (10.0 mmol/L) after meals, or above 130 mg/dL (7.2 mmol/L) when fasting. Common in people with diabetes whose insulin production or sensitivity is impaired.
Hypoglycemia	Low blood glucose. Generally defined as below 70 mg/dL (3.9 mmol/L). Can cause symptoms ranging from dizziness and confusion to seizures. Detecting hypoglycemic events quickly is one of the most critical functions of a CGM.
Dexcom G7	A seventh-generation continuous glucose monitor made by Dexcom. It uses a small filament inserted just under the skin and transmits glucose readings every 5 minutes to a smartphone or smartwatch via Bluetooth. Widely considered one of the most accurate consumer CGMs available.
Fine-tuning	In machine learning, the process of taking a model already trained on a broad dataset and continuing its training on a smaller, more specific dataset. Used here to adapt a general glucose model to an individual’s personal physiology.