A new research paper from Apple’s machine learning team suggests that the tech giant is looking at using artificial intelligence algorithms on behavioural data for future health functions on the Apple Watch wearable.
The article, published on the arxiv.org preprint server, discusses how this behavioural data – the way people move, exercise, and sleep – may be much more effective at spotting emerging health conditions than relying only on the current PPG sensor-based, physiological data included in wearable devices, such as heart rate or blood oxygen levels.
It is not certain that the new Wearable Behaviour Model (WBM) AI will make it into a finished product – not least because of the privacy considerations of having this type of health data recorded by a consumer tech company – but it provides an insight into possible future directions for health wearables.
The main advantage of WBM is that it draws on data generated over days and weeks, rather than the seconds-long timeframe for raw sensor data, which can be prone to short-term and transient factors affecting the individual, according to the authors.
The WBM was trained on 2.5 billion hours of wearable data from 162,000 people enrolled in the Apple Heart and Movement Study and put through its paces on 57 health detection tasks, ranging from identifying specific health conditions to detecting the use of common medicines like beta blockers and painkillers.
Layering WBM over regular sensor analysis was found to be significantly better at most of the tasks, including predicting atrial fibrillation (AFib), compared to PPG alone.
The combination was particularly effective at spotting pregnancy – 92% accurate, according to the authors – as well as detecting sleep disturbances, infections, and injury. The only exception was identifying diabetes, for which PPG alone was more effective.
“If developed and deployed safely and responsibly, predictive models built on wearable data like WBM hold significant promise for clinical impact,” the researchers write in the paper, pointing to potential applications in supporting clinical decision-making, triaging patients for follow-up, and supporting ‘just-in-time’ clinical interventions.
“By enabling continuous, non-invasive monitoring and early detection of meaningful health events, such models could support more proactive and personalised care – particularly for conditions where behavioural signals are strong early indicators.”