Can Your Smart Ring Detect Illness Before You Feel Symptoms
TL;DR
A March 2026 Nature study found that sleep and skin temperature data from consumer wearables can detect diabetes with 73% accuracy before clinical diagnosis. A separate January 2026 paper trained a multimodal sleep foundation model on 100,000+ nights of wearable data and found it could predict cardiovascular disease, sleep apnea, and diabetes from sleep patterns alone. These studies are the strongest evidence yet that the sensors in your smart ring can do more than track recovery. But there is a gap between what the research shows and what your wearable actually tells you, and it has nothing to do with the hardware.
In March 2026, a team from UC San Francisco and UC Berkeley published a paper in Nature Digital Medicine that should have gotten more attention than it did. Varun Viswanath, Ashley Mason, and Benjamin Lee Smarr took sleep and temperature data from consumer wearables and used it to screen for diabetes. Not continuous glucose monitors. Not finger-prick blood tests. Just the same PPG and temperature sensors that are in every Oura ring, every Whoop band, and every Fitbit.
The model hit 73% AUC for detecting diabetes. That is not diagnostic. But it is well above random, and it used data the device was already collecting for sleep tracking.
A second paper, published in Nature in January 2026 by Emmanuel Mignot's group at Stanford, trained a multimodal sleep foundation model on over 100,000 nights of wearable data. The model predicted cardiovascular disease risk, sleep apnea severity, and diabetes status from sleep architecture alone. Not from blood work. From the shape of your night.
These two papers represent a shift in what wearables can do. The question is whether the companies selling them will let you use it.
What the sensors actually capture
A smart ring collects roughly five streams of raw data: photoplethysmography (PPG) for heart rate and blood volume, an accelerometer for movement, a gyroscope for orientation, a temperature sensor for skin temperature, and in some cases a photodiode for SpO2. That is it. No chemical assays. No blood draws. No continuous glucose electrodes.
The diabetes detection study used two of these: skin temperature and sleep timing. The researchers found that people with undiagnosed diabetes showed a specific pattern: higher nocturnal skin temperature, more fragmented sleep, and a delayed temperature nadir. The temperature difference was small, about 0.3 to 0.5 degrees Celsius, but it was consistent across the 4,200 participants in the retrospective cohort.
The sleep foundation model used a broader set of features: heart rate variability, respiratory rate derived from the accelerometer, sleep stage timing, and movement patterns during sleep. It found that people with certain cardiovascular conditions spent less time in deep sleep and had higher heart rate variability during REM, even when their total sleep time looked normal.
Neither study required a sensor that does not already exist in a $160 smart ring.
The temperature sensor in a smart ring is a negative temperature coefficient thermistor. It measures the temperature of the skin on your finger, which is not the same as your core body temperature. The difference matters. Skin temperature lags core temperature by 15 to 30 minutes and is affected by ambient temperature, blood flow to the extremities, and whether your hand is under the blanket. The Viswanath study controlled for this by looking at the rate of temperature change overnight, not the absolute value. They found that the temperature nadir, the lowest point in the nightly temperature curve, occurred later in people with elevated blood glucose. The delay was about 45 minutes on average.
The PPG sensor in a smart ring uses green and red LEDs to measure blood volume changes in the finger. The green LED penetrates shallow tissue and picks up the pulsatile component of blood flow. The red LED goes deeper and can measure blood oxygen saturation. For the diabetes study, the researchers used heart rate variability derived from the PPG signal, not the raw optical data. HRV is a proxy for autonomic nervous system function, and autonomic dysfunction is an early sign of metabolic disease. The ring captures this data every night anyway, for sleep staging. The study just showed that the same data has diagnostic value the companies are not using.
The gap between research and product
Here is where it gets frustrating. The research exists. The sensors exist. The algorithms exist. But the products do not use them for this.
Oura has a feature called Illness Detection that flags when your temperature and HRV deviate from baseline. It is marketed as a way to catch early signs of sickness. In practice, it flags about 60% of COVID-19 cases before symptom onset, per Oura's own published data. That is useful. But Oura does not screen for diabetes, cardiovascular risk, or sleep apnea with the same data, even though the research shows it is possible.
Why not? Three reasons.
Regulatory. Screening for medical conditions turns a wellness device into a medical device. That means FDA clearance, clinical trials, and liability. Oura has not filed for FDA clearance on any disease screening feature. Neither has Whoop, Fitbit, or Garmin. The moment a wearable says you might have diabetes, it is making a medical claim, and the regulatory burden is significant.
Business model. Oura charges $5.99 per month for its subscription. The illness detection feature is part of that subscription. If Oura added diabetes screening, it would either need to charge more or justify the existing price. The current model works: give users enough insights to keep them subscribed, but not so many that the feature becomes a must-have that requires regulatory scrutiny.
Data architecture. Most wearables send your data to the cloud for analysis. That means the company has your raw biometric stream. If they wanted to run disease screening models, they could. But doing so would create a medical record, which changes the legal status of the data. HIPAA, GDPR, data breach notification laws, all of it applies differently when you are screening for disease versus tracking sleep.
Why on-device processing changes the calculation
This is where Pulsyn's architecture matters for this specific problem.
If disease screening runs on-device, the regulatory picture changes. The device is not transmitting diagnostic information to a company server. The model runs locally, on your phone, and the output stays on your phone. The company never sees the raw data and never makes a diagnosis. The device surfaces a pattern: your sleep and temperature look similar to patterns associated with early metabolic changes. That is a wellness insight, not a medical diagnosis, and it stays in the regulatory gray zone where consumer wearables already operate.
The privacy argument is not abstract here. If your wearable screens for diabetes, the result is health information. If that information lives on a company server, it is subject to subpoena, data breaches, and sale to data brokers. If it lives on your phone, encrypted with a key only you have, it is yours.
The latency argument matters too. Illness detection is time-sensitive. If your wearable detects a deviation from baseline at 3 AM, you want that information available when you wake up, not after the data has been uploaded, processed in a cloud data center, and sent back. On-device inference runs in milliseconds. Cloud inference adds minutes to hours, depending on connectivity.
What the research actually means for a smart ring user
Let me be specific about what these studies do and do not show.
The diabetes detection study was retrospective. The researchers had access to wearable data that was already collected, and they ran their model after the fact. That is different from a prospective study where the device flags people in real time and they go get tested. The 73% AUC is promising, but it is not ready for clinical use. The false positive rate would generate too many unnecessary doctor visits.
The sleep foundation model is more thoroughly validated. It was trained on a larger dataset and validated on multiple outcomes. But it was trained on data from wrist-worn devices, not rings. The signal quality is different. A ring has less motion artifact during sleep but a narrower PPG signal. The model would need retraining for the ring form factor.
What these studies do show is that the signal is there. The sensors in a smart ring capture enough information to detect physiological changes associated with disease. The question is whether the industry will build the models and ship them to users.
The GAN-based imputation paper from Barteit et al., published in Nature in March 2026, addresses a related problem: wearable data is full of gaps. People forget to charge their devices. They take them off for showers. The sensor loses contact during certain sleep positions. The GAN model fills in missing data with synthetic but statistically valid values, which improves the accuracy of downstream disease detection models. This matters because the diabetes study excluded participants with more than 30% missing data, which would rule out a significant portion of real-world users. If the imputation approach works at scale, it removes one of the biggest practical barriers to deploying these models on consumer devices.
The combination of these three papers, the diabetes detection study, the sleep foundation model, and the GAN imputation method, forms a technical stack that did not exist two years ago. The sensors, the models, and the data-cleaning pipeline are all mature enough for production use. The only missing piece is a company willing to ship it.
The companies that are actually doing this
A few companies are moving in this direction.
Fitbit received FDA clearance for its atrial fibrillation detection algorithm in 2024. That is a specific, narrow use case: detect AFib from PPG data. It required a clinical trial with over 450,000 participants. Fitbit can do this because Google has the resources to get through FDA clearance and the legal infrastructure to handle the liability.
Apple Watch has FDA-cleared ECG and AFib detection. Apple also published a study in 2025 showing that wrist temperature data could predict ovulation with 94% accuracy. That is not a disease screen, but it is a physiological state detection that uses the same sensors.
Oura has the illness detection feature but has not pursued FDA clearance for any specific condition. Oura's published research on COVID-19 detection is the closest the ring form factor has come to disease screening, and it works reasonably well. But Oura has not extended it to other conditions.
Nobody in the smart ring space is doing what the Nature papers describe: using sleep and temperature data to screen for metabolic and cardiovascular disease. The ring form factor is actually better suited for this than the wrist, because the finger has better perfusion for PPG and the ring has less motion artifact during sleep. The hardware is ready. The software is not.
What Pulsyn is doing about it
I am building Pulsyn with on-device AI from day one. That means the model that analyzes your sleep, temperature, and HRV runs on your phone, not in a cloud server. The architecture supports running inference locally, which means we can add pattern detection features without changing the privacy model.
I am not going to claim that Pulsyn will detect diabetes at launch. It will not. The research is promising but not product-ready. What Pulsyn will do is surface deviations from your baseline and let you decide what to do with them. If your skin temperature runs 0.4 degrees above your 30-day average for three consecutive nights, and your HRV drops by 15%, the device will tell you. It will not tell you what it means. That is for you and your doctor.
The reason I am building it this way is that the alternative, sending your biometric data to a cloud server for analysis, creates a permanent record of your physiological state that you cannot control. If that data shows patterns consistent with early diabetes, and the company stores it, that information exists outside your control. It can be subpoenaed. It can be sold. It can be breached.
On-device processing means the pattern detection happens and the result stays on your phone. The company never sees the raw data. The model never phones home. If you want to share the output with your doctor, you can export it. If you do not, it stays encrypted on your device.
The open question
The Nature papers from 2026 show that consumer wearables can detect disease from sleep and temperature data. The sensors are already in the devices people wear every night. The algorithms exist in academic literature. The gap is not technical. It is regulatory and business-model.
The question is whether any wearable company will ship a product that uses these sensors for what they are actually capable of, without requiring a subscription, without uploading your data to a cloud server, and without turning your health information into a corporate asset.
I think the answer is yes, eventually. I am building Pulsyn to be that product. But I also think the industry is moving too slowly, and the people who would benefit from early detection are the ones paying the price.
About the author
James Hoffmann is the founder of Pulsyn. He has been building on-device AI systems for health wearables for two years.
References
Viswanath, V. K., Mason, A. E., & Smarr, B. L. (2026). Sleep and temperature data from wearable devices support noninvasive detection of diabetes mellitus in a large-scale, retrospective analysis. Nature Digital Medicine.
Mignot, E., Brink-Kjaer, A., & Kjaer, M. (2026). A multimodal sleep foundation model for disease prediction. Nature.
Barteit, S., Dendorfer, A., & Obor, D. (2026). Overcoming Data Loss in Wearable Disease Detection with GAN-Based Imputation. Nature Digital Medicine.
Oura Health. (2025). Oura Ring illness detection: Early COVID-19 symptom detection using wearable-derived physiological data. Oura Research.
Fitbit / Google. (2024). FDA clearance of PPG-based atrial fibrillation detection algorithm. FDA 510(k) clearance.
