📄 health informatics

A foundation model of wearable pulse oximetry reveals physiological signatures of health and cardiometabolic risk

The paper introduces PulseOx-FM, a self-supervised foundation model trained on millions of wearable pulse oximetry segments that outperforms existing methods in predicting diverse cardiometabolic and neuropsychiatric health risks, including future hypertension and next-day glycemic states, thereby establishing a powerful non-invasive tool for global health risk stratification.

Original authors: Kohn, S., Lutsker, G., Diament, A., Shilo, S., Gabet, A., Sasson, G., Wolf, G., Wolf, A., Godneva, A., Weinberger, A., Rossman, H., Segal, E.

Published 2026-07-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Kohn, S., Lutsker, G., Diament, A., Shilo, S., Gabet, A., Sasson, G., Wolf, G., Wolf, A., Godneva, A., Weinberger, A., Rossman, H., Segal, E.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your body is a complex orchestra playing a continuous, silent symphony. Even when you are asleep and still, your heart, blood vessels, and nervous system are constantly adjusting the tempo, volume, and rhythm of this music. For years, doctors have only been able to hear the loudest instruments: the beat of the heart and the level of oxygen. But there is a whole hidden layer of subtle, high-resolution music in the way your blood flows that has been largely ignored.

This paper introduces PulseOx-FM, a new "musical translator" (a foundation model) that learns to read this hidden layer of music using data from standard wearable pulse oximeters (the little clip or watch that measures heart rate and oxygen).

Here is a breakdown of what the researchers did and found, using simple analogies:

1. The "Super-Listener" Training

The researchers didn't teach the AI by giving it a textbook of rules. Instead, they used a method called "self-supervised learning."

The Analogy: Imagine giving a student 7 million short audio clips of a river flowing, but never telling them what a river is. The student has to listen, find patterns, and learn to predict what comes next in the flow. Eventually, the student understands the river's nature so well that they can describe the river's health, speed, and even the type of rocks underneath just by listening to the water.
The Reality: The AI (PulseOx-FM) was trained on 6.9 million segments of sleep data from over 10,000 people. It learned to reconstruct the raw blood flow waveforms on its own, discovering deep patterns that human experts had missed.

2. The "Biological Age" Test

To see if the AI actually learned anything useful, the researchers asked it to guess a person's age.

The Analogy: Think of your blood vessels like rubber bands. As you get older, rubber bands get stiff and lose their bounce. Old ways of measuring this were like looking at a rubber band with a ruler (measuring specific points). The new AI is like a master craftsman who can pick up the rubber band, feel its tension, and instantly know exactly how old the rubber band is, even if it's been stretched in weird ways.
The Result: The AI guessed people's ages much more accurately than existing methods. Crucially, it did this so well that it could even guess the age of patients in a completely different setting (surgery under anesthesia) without needing to be retrained. This proves it learned the fundamental physics of human blood flow, not just the specific "sound" of sleeping people.

3. The "Health Snapshot"

The researchers tested if the AI could tell you about a person's current health just by listening to their sleep pulse.

The Analogy: If you walk into a house and hear the creaking of the floorboards, the hum of the fridge, and the draft under the door, you can guess if the house is old, if the roof is leaking, or if the pipes are rusty. You don't need to see the owner to know their health.
The Result: The AI could predict a wide range of health issues, including:
- Heart and Blood Pressure: It was very good at spotting high blood pressure and signs of heart valve issues.
- Sleep Apnea: It could detect breathing interruptions during sleep better than standard sleep trackers.
- Mental Health: Surprisingly, it could also detect signs of anxiety and depression, likely because these conditions change how the nervous system controls the heart.
- Medication: It could tell if someone was taking specific heart or blood pressure medications, acting like a "drug detector" based on how the drugs changed the blood flow rhythm.

4. The "Tomorrow's Weather" Forecast

One of the most exciting findings was about predicting what happens the next day.

The Analogy: Usually, weather apps look at today's clouds to guess tomorrow's rain. This AI is like a farmer who looks at the soil moisture and the wind direction tonight to predict exactly how thirsty the crops will be tomorrow morning, even before the sun comes up.
The Result: The AI could predict next-day blood sugar levels, energy expenditure (how many calories you burn), and even what you might eat, based only on your sleep pulse.
The Big Discovery: The researchers wanted to know: "Is the AI just guessing what you ate for breakfast?" They used a special "detective" system (an AI agent) to prove that the signal was direct. The sleep pulse wasn't just a shadow of your diet; it was a direct physiological signal from your body telling you how your blood sugar would behave the next day, regardless of what you ate.

5. The "Universal Translator"

Finally, the paper shows that this tool works across different "dialects" of physiology.

The Analogy: Most tools are like a dictionary that only translates English to French. PulseOx-FM is like a universal translator that learned English (sleeping people) but can instantly understand and translate Japanese (surgical patients under anesthesia) without needing a new dictionary.
The Result: The model trained on healthy people sleeping at home worked just as well on sick people in an operating room. This suggests it has learned the "universal grammar" of human blood flow.

Summary

In short, the paper claims that by using a massive AI to listen to the subtle, high-resolution "music" of blood flow during sleep, we can now:

Measure biological age more accurately than ever before.
Detect hidden health risks (like heart disease, sleep apnea, and anxiety) without invasive tests.
Predict next-day metabolic states (like blood sugar) directly from the night's data.

The authors emphasize that this is a foundation model—a powerful engine that has learned the underlying rules of physiology. They state that this opens the door to using simple, cheap wearable devices to monitor global health risks in a way that was previously only possible with expensive, complex hospital equipment. However, they also note that while the tool is powerful, it currently mixes the signals of disease with the signals of medication, and more testing is needed before it can be used as a standard medical diagnostic tool.

Technical Summary: A Foundation Model of Wearable Pulse Oximetry

Problem Statement
Photoplethysmography (PPG) is a ubiquitous, non-invasive optical technology used in consumer wearables and clinical devices to monitor heart rate and oxygen saturation. While the raw PPG waveform contains rich physiological data encoding cardiovascular, autonomic, and respiratory states, this information remains largely uninterpreted. Current deep learning applications for PPG are constrained by task-specific supervision (e.g., sleep staging) or rely on hand-crafted features (e.g., fiducial points) that are often overfitted to specific diagnostic targets like the apnea-hypopnea index (AHI). Consequently, there is a technological divide: scalable PPG models lack broad predictive power, while comprehensive physiological models (e.g., those using full polysomnography) are difficult to scale for population health. The authors posit that a foundation model trained on large-scale, unlabeled PPG data could unlock intrinsic physiological signatures for global health risk stratification.

Methodology
The study introduces PulseOx-FM, a transformer-based foundation model designed to learn latent representations from raw peripheral pulse-oximetry signals.

Data Source: The model was pretrained on the Human Phenotype Project (HPP), a large-scale longitudinal study involving 11,311 participants. The dataset comprises 45,068 overnight sleep monitoring recordings (310,264 hours) collected via the WatchPAT device, which captures SpO2, pulse rate, and peripheral blood flow waveforms (via Peripheral Arterial Tonometry, PAT).
Pretraining: PulseOx-FM utilizes a Masked Auto-Encoder (MAE) objective. It was trained on 6,995,558 segments of 2-minute pulse oximetry data. This self-supervised approach allows the model to learn physiological representations without manual labels by reconstructing masked portions of the input signal.
Architecture: The model employs a transformer encoder-decoder structure to generate 256-dimensional latent embeddings. These embeddings are pooled across multiple segments per night and fed into lightweight, task-specific prediction heads.
Evaluation Strategy:
- Benchmarking: Performance was compared against established baselines, including open-source feature extraction (PyPPG) and proprietary device algorithms (WatchPAT).
- Downstream Tasks: The frozen embeddings were evaluated on 64 phenotypic targets, including continuous measures (blood tests, body composition, sleep parameters) and binary outcomes (diagnoses, medication intake).
- Generalization: The model was tested on an external, out-of-distribution cohort (VitalDB) comprising 6,156 participants undergoing intraoperative monitoring under general anesthesia, a context vastly different from the training data.
- Longitudinal Analysis: The study analyzed within-person stability over time (consecutive nights vs. two years apart) and the ability to predict next-day metabolic states (glucose, diet, activity) while controlling for sleep architecture and dietary intake.

Key Results

Age Prediction as a Health Benchmark:
- PulseOx-FM significantly outperformed both PyPPG (hand-crafted features) and WatchPAT proprietary features in predicting chronological age, a proxy for global health.
- On the HPP test set, PulseOx-FM achieved a Pearson correlation ( $r$ ) of 0.69, compared to 0.65 for PyPPG and 0.64 for WatchPAT.
- Crucially, the model maintained robust performance on the external VitalDB surgical cohort ( $r=0.70$ ), demonstrating that the learned representations generalize from ambulatory sleep to acute intraoperative states without fine-tuning.
Broad Phenotypic Prediction:
- Adding PulseOx-FM embeddings to demographic baselines (age, sex, BMI) significantly improved predictions for 64 targets.
- Cardiometabolic Domain: The largest gains were observed in respiratory disturbance index ( $r: 0.36 \to 0.76$ ), AHI ( $0.43 \to 0.70$ ), systolic/diastolic blood pressure, triglycerides, and carotid intima-media thickness.
- Medication Detection: The model showed substantial improvements in detecting medication classes, particularly psychoanaleptics (AUC $0.49 \to 0.70$ ) and antihypertensives (beta-blockers, RAAS inhibitors, calcium channel blockers), suggesting the embeddings capture pharmacodynamic signatures in the vasculature.
- Sleep Architecture: The model recovered sleep parameters (e.g., sleep efficiency, total wake time) that demographics alone could not predict, indicating it internalizes individual-level sleep structure from raw waveforms.
Longitudinal and Next-Day Predictions:
- Hypertension Risk: The model prospectively identified the two-year incidence of hypertension in normotensive individuals (AUC $0.51 \to 0.64$ ), suggesting it captures pre-clinical vascular signatures.
- Within-Person Stability: Embeddings for the same individual on consecutive nights were significantly closer (median L2 distance 0.68) than those of demographically matched strangers (1.55), confirming a stable, individual-specific physiological fingerprint.
- Next-Day Metabolic State: Embeddings significantly improved the prediction of next-day glucose management indicator (GMI), energy expenditure, and activity levels compared to sleep-architecture features alone.
- Diet-Independent Glycemic Signal: Using a literature-driven agentic analysis system, the authors performed causal mediation analysis. They found that the association between nightly embeddings and next-day glycemia was predominantly a direct physiological effect (approx. 100% of the total effect), not mediated by next-day dietary intake.

Significance and Claims
The paper claims that PulseOx-FM establishes a generalizable framework for encoding physiological patterns from sleep using only a single, scalable optical signal. Its significance lies in bridging the gap between the accessibility of consumer wearables and the predictive depth of comprehensive physiological modeling.

Intrinsic Physiological Learning: The model's ability to transfer from sleep to anesthesia without fine-tuning suggests it learns intrinsic, systemic properties of human vascular physiology rather than dataset-specific artifacts.
Beyond Vital Signs: The findings demonstrate that raw PPG waveforms contain information extending far beyond heart rate and SpO2, capturing vascular aging, cardiometabolic risk, and autonomic modulation.
Precision Medicine Potential: By identifying sub-clinical vascular signals (e.g., future hypertension risk) and diet-independent metabolic states, PulseOx-FM offers a pathway toward passive, continuous, and accessible longitudinal health monitoring for global risk stratification.

The authors conclude that while the model shows promise for clinical translation, prospective validation in diverse populations and further work to disentangle disease signatures from pharmacological effects are necessary before clinical deployment.