Generative AI in digital health: turning data into action

What are generative insights?

Generative insights are machine-synthesised interpretations of multi-sensor data that answer human questions such as “Am I recovering?”, “Am I getting ill?” or “What should I do next?”, rather than simply plotting and presenting numbers and infographics with little-to-no actionable recommendations for the user. Generative insights typically combine long-term baseline measurements with contextual factors such as sleep, activity and environment to output summaries, risk scores and suggested actions for the user.

The potential of generative AI was spotlighted at a recent conference, “The future of medical-grade wearables: From consumer to clinic” at The Royal Society of Medicine in London. A recurring message arose – that wearable technology powered by AI can promise more accurate, less biased and more meaningful health data, but only if it is designed to fit real lives rather than idealised routines. In this framing, the value of biomarker data goes beyond better tracking of health insights, to reducing the guesswork in conversations between patients and healthcare professionals (HCPs), especially when trying to answer the difficult retrospective question: “How have you actually been over the last six months?”

Clinical practice applications account for approximately 57% of the global digital biomarkers market share, driven largely by an increasing integration of digital monitoring technologies into hospitals, clinics and telemedicine platforms. Additionally, the rise of AI and machine learning (ML) integrated in health tools has resulted in regulatory agencies already treating many such algorithms as Software as a Medical Device (SaMD) when used for diagnosing, predicting, or treating conditions. With the digital biomarkers market projected to reach $32 billion by 2034 and the refinement of AI models, we will continue to see innovations develop in predictive and personalised treatments.

Current applications of generative insights

Many digital health products have moved away from showing passive dashboards of a user’s health stats to presenting timely, generative insights based on their data. Biomarker data, such as heart rate, respiratory rate, temperature, heart rate variability (HRV) and sleep, is captured and measured using off-the shelf wearable technology and synthesised into personalised narratives, predictions and recommendations.

Around 50% of the world’s population now use wearable devices to monitor their health, reflecting a broader shift in attitudes towards healthcare – from reactive treatment to proactive prevention. Many industry leaders in the wearable space such as Google, Garmin, Oura and WHOOP are harnessing AI models to deliver highly personalised health and wellness monitoring experiences through generative insights.

Beyond wearables, generative AI insights are supporting clinical research and delivery of care – informing decentralised trials, enabling earlier detection of infection and powering real-time interventions.

Fitbit integrated with Google Health – personalised coaching built on continuous health and lifestyle data

Google is integrating Fitbit into a broader Google Health experience, with a Gemini-powered Health Coach for Fitbit Premium users. The Health Coach can help log symptoms and generate more personalised plans for fitness, sleep and nutrition using health metrics, environmental context and linked medical records. This turns passive tracking into more adaptive, context-aware support.

WHOOP wearable – just-in-time interventions (JITAIs)

WHOOP’s latest updates show a similar move to Fitbit, combining clinician access with more contextual AI features such as “My Memory”. The My Memory feature lets users refine the personal goals and routines that shape interpretation of their data, while proactive check-ins turn biometrics and habits into timely recommendations. JITAIs, such as WHOOP’s new proactive “check-ins” feature, tailor content and timing to user receptivity, context and momentary need. This makes users more likely to act on prompts in the moment and reduces notification fatigue compared to one-size-fits-all prompts.

Wearables and passive monitoring used in clinical trials

In decentralised trials, wearables and passive monitoring can reduce the need for participants to travel to sites or manually enter data, allowing information to be captured more naturally and in the context of participants’ lives. This can make trials more inclusive and less burdensome, while also providing researchers with richer longitudinal datasets. These systems may help researchers spot subtle physiological changes earlier and build a more realistic picture of health over time.

Of course, there are potential challenges in this area too. In the case of clinical trials, collecting data using AI-powered wearables poses re-identification risks, where algorithms can re-identify anonymised biometric and health data. Additionally, rules and protections would need to be in place so no data would be sold to third parties (compared to what is typically done).

Hybrid closed-loop system – automated insulin delivery

In diabetes care, hybrid-closed loop insulin delivery systems, or ‘artificial pancreases’, continuously translate glucose data and contextual inputs into safe insulin micro-adjustments. This is done through a continuous glucose monitor (CGM) connected to an insulin pump via a smart algorithm. Randomised trials, real‑world studies and policy analyses have shown these systems offer improved time‑in‑range and quality of life. This is the strongest proof that data-driven insights can become actions at scale.

Earlier detection of infection

Wearables can detect deviations from a user’s baseline. This data can flag respiratory infections days before symptom onset. Wearables use algorithms to both track biomarker data progression as well as notify users when a deviation is noticed, for example with the Oura ring’s ‘signs of strain’ app feature. Presenting this information to users can help set in motion the appropriate actions to prevent any further strain or development of illness. These features can enable users to have greater visibility and control of their health. Additionally, Oura has recently filed for an IPO to formally position itself as a health intelligence and preventative health platform, rather than just a wearable company. This is an example of how fitness and wellness companies are moving into the healthcare space.

With all of the above scenarios, there are challenges around privacy, accuracy and regulation, particularly when AI is used, for example, to identify subtle patterns to propose novel endpoints or to power conversational agents.

Generative AI decision-making in healthcare – when data overtakes human judgement

As the potential of GenAI to support patients expands, health monitoring tools can move beyond indicating how someone is feeling and can start to shape it. This is where placebo and nocebo effects become important. The meaning attached to a score, alert or recommendation can influence symptoms, behaviour and decision-making. Sometimes this anticipates needs, other times it leads to harm.

Defining the placebo versus nocebo effect

The placebo effect can be defined as any improvement of illnesses or reduction of subjective symptoms due to someone believing a treatment, message or signal has helped despite no physical intervention.

For example, in healthcare studies, false feedback of a slow heart rate has been shown to reduce anxiety in a lab setting – a powerful demonstration that perceived bio-signals shape felt experience.

‘Nocebo’ is also a powerful psychological effect, referring to undesirable symptoms or illnesses that follow interventions also lacking known physical effects. Negative expectations can trigger fear and anxiety which in turn generate a nocebo effect. There is evidence that informing someone about painful interventions increases the intensity of perceived pain.

For example, if an app connected to a wearable says your “readiness” or “battery” is low, you might then feel worse and perform worse, despite the fact that before absorbing this data your body feels fine.

The term “orthosomnia” is a type of nocebo effect which describes fixation on sleep metrics that ironically worsen sleep. Case reports and subsequent public-health explanations document patients stretching their amount of time in bed to “improve numbers,” escalating anxiety and insomnia.

Another example of the nocebo effect relates to anxiety amplification in heart rhythm monitoring (HRM). For people with atrial fibrillation, consumer alerts can increase health-related anxiety, clinic call frequency and diagnostic testing orders – even when clinical need is unclear. A Journal of the American Heart Association study reports ~1 in 5 users experiencing intense fear in response to irregular rhythm notifications.

Left unchecked, generative insights can override how a person feels, provoke reassurance-seeking and shift agency away from the user to the AI-enabled device.

The importance of context and regulation

Some measurements taken from wearables, such as respiratory rate, can be accurate at rest and during sleep, but still ride on assumptions and noise. Algorithms often perform best on the population they were trained on. It is key that users are informed of the transparencies behind the technology they use, so that it elevates their health, rather than dictates it.

When developing such wearables, some design considerations could be to present ranges, uncertainty bands and confidence-in-context, e.g., “high confidence during sleep, lower confidence during post-workout” to allow a user’s subjective state to co-author the narrative.

How can good design optimise the user experience of generative AI-powered wearable devices?

Good UX design can help generative AI-enabled wearable devices present insights in ways that build trust, reduce anxiety and empower user’s own feelings, rather than treating data as the only ‘source of truth’. Below are key UX design considerations for these technologies.

1. User first, data second

Lead with the user’s self-evaluation, e.g. “you feel good today”, then layer the data, e.g. “last night’s respiratory rate was slightly down”. This prevents data priming – where data can influence a user’s perception of their own health – from overwriting the user’s felt experience.

2. Be transparent around data collection

Present a confidence/strength of evidence badge when presenting information back to users, e.g., “High confidence of REM sleep, medium confidence of respiratory rate throughout day due to increased movement”. This can help users to understand how reliable an insight is to allow them to make informed decisions on whether to take action based on it.

3. Focus on data trends rather than single-day readings

Show a user their health data over time to build a greater picture of biomarker progression. Offer different ways of presenting insights for people with metric fixation or anxiety. Users will be able to identify patterns and changes over time to help them manage reaction to variation – the overview is more balanced rather than isolated fluctuations which may cause the user to overreact.

4. Psychoeducation

Include resources where users can be aware of the placebo and nocebo effect when tracking their health data. Explain how the brain chemistry and psychology can alter physical symptoms, such as how reading “worse” health data can make you feel worse even if this may not be true. Interoception check-ins could help users ensure they prioritise their own feelings over their tracked data.

5. Allow users to control the levels of automation

Provide users with the option of having semi-automatic changes (ask to confirm) or fully automatic when it comes to adjusting their health plan and always log each automated action with easy ‘un-do’ abilities. For example, in the fitness space, poor sleep data could suggest changing or removing a workout that day. The user should always have the final say on any plan changes.

From hints to actions: example scenarios of what’s next for GenAI tools

As GenAI-enabled health tools mature, they will move beyond nudging behaviour to taking safe, reversible actions on a user’s behalf, and should remain transparent about uncertainty and leaving the final say with the user. Wearables which work with AI-powered systems can work across both the healthcare and wellness space to function as assistants to optimise recovery and elevate patient care.

A near-future day might look like this:

Morning

Sam wakes with a stiff shoulder as a result of a recent rotator cuff injury. Overnight, their wearable has spotted signs of strain from their elevated HRV and poorer sleep and prompts to adjust the day’s plan: medication first, then a shorter physio session timed to when pain relief is strongest. A quick message explains it simply: Today’s a lighter rehab day to keep recovery moving.

Afternoon

As Sam works at their desk, the wearable picks up reduced movement and rising stress through a higher heart rate when typing at their computer. Instead of just tracking this, the system acts to prompt a short break, a heat pack and two targeted exercises. The system times each nudge around both medication effect and physio goals, keeping treatment joined up.

Evening

By evening, Sam’s pain is steadier and their movement is slightly better than usual. The system responds by scaling back the rest of the day, updating tomorrow’s rehab plan, and sending a clear summary: Medication timing was effective today and mobility has improved. Tomorrow’s plan has been adjusted to build on today’s progress.

Systems which are thoughtfully designed to quietly and seamlessly integrate to a user’s lifestyle whilst being powered by intelligent algorithms could help users understand and trust in their treatment plan to regain their health.

Conclusion

Generative insights have the potential to move digital health from passive tracking to timely, meaningful support. When grounded in context, uncertainty and user experience, they can help people act earlier, reduce decision fatigue and support better conversations with healthcare professionals.

Insights that overstate certainty or ignore how people feel risk creating anxiety and misplaced trust. Systems that respect interoception, explain limitations and keep users in control are far more likely to deliver benefit. The next step for digital health is clear: turning data into action in ways that are careful, reversible and user centred.