The idea that a watch could help detect health problems before they appear, and even to anticipate pregnanciesIt no longer sounds so far-fetched. A new study has shown how the combination of artificial intelligence and data passively collected by the Apple Watch can bring preventative medicine to the wrist, without the need for constant clinical trials.
To achieve this, a group of researchers has resorted to a very specific formula: using advanced AI algorithms capable of squeeze incomplete and gap-filled recordsThis is precisely the kind of information that smartwatches generate in real life. Instead of striving for perfect control, the model learns to work with what's available, just as the device is used on a daily basis.
A macro-study with 16.522 users and three million days of real data
The project relies on a massive database, made up of 16.522 people who used an Apple Watch over long periods of time. If all those records are added together, the total is approximately three million days of actual use, a figure that allows us to observe patterns that would be impossible to detect on a small scale.
Each user generated up to 63 different metrics related to your healthThe data was organized into several sections: cardiovascular parameters (such as heart rate and heart rate variability), respiratory indicators, sleep quality, physical activity level, and other general statistics on daily routines. Not all participants used the watch with the same consistency, but this was not considered an insurmountable problem.
In fact, a key part of the study is that Most of the data was not accompanied by a clear medical diagnosis.Only around 15% of participants had a labeled medical history, that is, with formally recorded diseases that could serve as a direct reference for the algorithm.
Instead of discarding all that unlabeled data, the team opted for a strategy of self-learning or self-monitoringFirst, the model was trained using the complete set of records, without needing to know which person had which disease. Later, it was refined with the small group of users who did have confirmed diagnoses.
This approach allows to take advantage of data that, in other contexts, would be considered too noisy or irregularThe model not only looks at specific values, but also tries to understand how each individual behaves over time, including their ups and downs, days without a clock, and changes in routine.
An AI that understands the gaps in Apple Watch data
The technical core of the system is inspired by architecture JEPA (Joint Embedding Predictive Architecture)This type of model is designed to understand broader contexts rather than predict isolated data points. Unlike language models that attempt to guess the next word, the priority here is to construct a coherent representation of the person's overall state.
To do this, each observation recorded by the watch is transformed into a kind of token that includes the day, the type of metric, and the measured valueA masking mechanism is applied to this sequence: certain parts of the information are deliberately hidden so that the AI ​​has to infer what might be in those gaps.
The key is that the model doesn't try to fill in the empty spaces with an exact number, but interpret what that absence means within each user's patternFor example, several days without sleep records are not automatically considered a mistake, but rather another element of behavior that can provide context about habits or specific changes.
This is especially useful when working with wbecause its use is far from perfect: The watch stays charging on the nightstand, the battery runs out by mid-afternoon, or some sensor malfunctions intermittently.The model, dubbed JETS by the researchers, is specifically designed to take advantage of this chaos and remain useful with highly irregular data.
The study's authors point out that some metrics were only available around a 0,4% of total time For some users, registrations were frequent, while others were almost daily. Even so, the AI ​​managed to extract relevant patterns that would likely have gone unnoticed with more traditional approaches, focused on ordered and complete series.
Ability to anticipate hypertension, sleep apnea, and other health risks
Once trained, the model underwent testing with various specific medical conditionsThese conditions are closely linked to cardiovascular health and sleep disorders. They include high blood pressure, sick sinus syndrome, various types of chronic fatigue, and problems consistent with sleep apnea episodes.
In the case of the hypertensionThe study data points to a discrimination capacity of around 86,8% between people with and without this condition. It doesn't mean it gets the diagnosis right down to the millimeter in every case, but it is quite effective at separating those who probably have a problem from those who are less likely to.
The model also showed good performance in detecting patterns linked to sick breast syndrome and chronic fatigue syndromeCompared to other algorithms and comparative methods, it wasn't always number one in every metric, but it did show a consistent advantage when it came to working with incomplete and highly disparate records.
The metrics used to evaluate the system do not directly count "successes," but rather measure how well it prioritizes cases with the highest probability of riskThis form of evaluation fits with preventive medicine, where what is really important is deciding who should be checked first, rather than nailing the definitive diagnosis at first glance.
In practice, the potential of this type of AI lies in act as a passive screening systemIt runs in the background and alerts users who should consult a healthcare professional. The Apple Watch thus becomes an initial filter, using seemingly routine measurements to raise concerns when it detects something unusual.
Imperfect data that can help save lives
One of the most interesting ideas that emerges from the work is that Imperfect data is not synonymous with useless data.If analyzed with the right approach, they can be tremendously valuable, especially when collected over months or years in everyday life situations, far from the controlled environments of a laboratory.
Even very sporadic records can contribute to building a robust model of each person's overall health. By combining Information on physical activity, sleep quality, and heart rate behaviorThis yields a fairly comprehensive map that can reveal underlying problems that have not yet shown clear symptoms.
This approach reinforces the impression that Smartwatches like the Apple Watch can play an increasing role in continuous health monitoringIt is not essential to wear the device 24 hours a day or to obsess over pressing all the buttons: the key is to have systems capable of interpreting what is actually being measured with sound judgment.
The model developed by the researchers is designed precisely for that purpose: accompany the user in their normal routine without demanding perfect useFrom the accumulated data fragments, the AI ​​builds a kind of "dynamic portrait" of the person, stable enough to detect striking deviations.
For European healthcare systems, typically strained by high demand for care, tools of this kind could be a valuable addition to support clinical work. An algorithm that help prioritize cases or identify those who need a more urgent review It can be very useful, provided it is integrated judiciously and not used as a substitute for medical consultation.
Limitations, clinical validation, and data privacy
The authors of the work themselves insist that All of this still belongs to the realm of research.Just because a model works well in a study doesn't mean it's ready to become a commercial feature of the watch or to issue diagnoses on its own.
For now, it has not been conclusively proven how would this system behave in real clinical settingswhere very varied factors influence: differences between countries and populations, different ways of using the watch, lifestyle changes, medical treatments that alter the metrics or simply times when the person stops using the device.
Even with high rates of discrimination, AI is far from infallibleFalse positives can occur, generating unnecessary alarm, as can false negatives that overlook a significant risk. Therefore, any signal emitted by these types of models should be understood as a recommendation to seek professional advice, not as a definitive conclusion.
In this context, the role of Medical personnel remain absolutely centralPhysical examination, specific diagnostic tests, and a comprehensive patient assessment are elements that an algorithm fed solely with passive clock data cannot replicate. AI can help focus care more effectively, but the final decision must rest with the professionals.
To all this must be added the question of the privacy and protection of health dataThis is especially sensitive in Europe. The continuous analysis of such intimate information as heart rate, sleep patterns, or activity levels requires ensuring high levels of security and strictly complying with regulations such as the General Data Protection Regulation (GDPR).
Potential impact on preventive medicine in Spain and Europe
Beyond the precautions, the work points to a change of focus in the way we understand the Preventive medicine in Spain and the rest of EuropeMoving from occasional check-ups to almost constant monitoring, even if passive, opens the door to treating health problems at much earlier stages.
The Apple Watch already offers features such as abnormal heart rhythm alerts, pulse irregularity notifications, sleep logs, or fall detectionThis new type of model suggests a further step: that the device can collaborate in the detection of pathologies such as hypertension or certain forms of sleep apnea with a more global view of the user's behavior.
In the European context, where telemedicine and remote monitoring initiatives are growing, these advances would fit into projects of remote monitoring of chronic patients or people with risk factorsIntegrating data from the clock into the electronic health record, always with appropriate consent, would allow professionals to have a much more detailed picture between visits.
For the Spanish healthcare system, accustomed to managing waiting lists and high demand in primary care, having automated screening tools based on real data could to help refer earlier those who need more specific testsHowever, this would need to be accompanied by clear protocols, training for healthcare workers, and a rigorous evaluation of benefits and potential side effects, such as an overload of alerts.
The health ecosystem on the iPhone is also moving in that direction. The Health app has evolved from a simple repository to A dashboard that highlights trends, displays striking changes, and facilitates report exportsIf this type of AI model ends up being integrated into that environment, the European user could have a much more proactive assistant, but also one that is more demanding in terms of transparency and control of their data.
Everything suggests that we are at the beginning of a new stage in which an artificial intelligence trained with millions of days of Apple Watch use It acts as a silent radar for anomalies, while physicians retain the final say. The study's results show that even irregular recordings can offer valuable clues about hypertension, sleep apnea, or other problems, provided they are handled with scientific rigor and under the safeguards of European data protection regulations.
