AI Population Health Analytics: Using Wearable Data to Understand Risk at Scale

Population health has always been about aggregates. Which cohorts are at risk, which interventions reach the right members, where costs are heading and why. The challenge has never been the question. Most of the time, we lack the data available to answer it.

Traditional population health models are built on claims, lab results, and diagnostic codes. These are valuable, but they share a structural problem: they are episodic. A claims record tells you what happened at the point of care. A lab result tells you what a biomarker read on the morning it was drawn. Neither tells you what was happening in the six months between those events, which is often when the trajectory that led to a costly outcome was already visible.

Wearable data changes that. Not because it replaces clinical data, but because it fills the space between clinical touchpoints with continuous signal. When that signal is normalized, aggregated, and processed at population scale with AI, the picture of risk that emerges is fundamentally different from what claims alone can produce. That’s how we built our Population Health Dashboard. 

What Is The Problem with Traditional Population Health Data

The lag in traditional population health data is not a minor inconvenience. It is a structural constraint that determines what kinds of interventions are possible.

Claims data typically reaches insurers and health systems weeks to months after the event it describes. By the time a claims-based risk model flags a member as high risk, that member has often already incurred the cost the flag was meant to prevent. Lab results are ordered at clinical visits, which for many members happen once or twice a year. Diagnostic codes reflect what was documented at a point of contact, not what is developing in between.

Self-reported data, like questionnaires, symptom logs, patient-entered outcomes, is subject to recall bias, completion rates, and the fundamental problem that people under-report and under-track precisely when their health is declining and engagement drops.

The result is a population model that is always looking backward. Risk scores are computed from events that have already occurred. Cohorts are defined by diagnoses that have already been made. The intervention window, if there is one, is narrower than it should be.

What Wearable Data Adds to Health Analytics

Wearable devices generate continuous physiological and behavioral signals between clinical touchpoints. That continuity is what makes them valuable for population health at a level that goes beyond what individual wellness apps deliver.

Several specific signals matter at the population layer:

• Continuous baseline data. Resting heart rate, HRV, sleep duration and staging, and activity levels collected daily over months and years establish an individual baseline for each member. Deviations from that baseline are more predictive than comparisons to population averages, because they reflect something changing for that specific person.
• Early deviation detection. Wearable signals often shift weeks before a clinical event. Declining HRV, disrupted sleep, and reduced activity in combination are associated with oncoming illness, stress responses, and deteriorating chronic disease management, frequently before a symptom is reported or a clinical visit occurs.
• Behavioral patterns. Activity levels, sedentary time, sleep consistency, and recovery trends carry information about lifestyle factors that claims data cannot see. For chronic disease management and prevention programs, these patterns are as relevant as biomarkers.
• Longitudinal continuity. A wearable dataset covering twelve months of daily readings provides a fundamentally different analytical surface than a dataset of four lab results and two clinical visits over the same period.

The value is not in any single reading. It is in the pattern across time, and what that pattern reveals when analyzed at the scale of a whole member population.

Scaling From Individual Signals to Population-Level Intelligence

Moving from individual wearable data to population-level intelligence requires a set of analytical steps that go beyond what device apps or individual health dashboards provide.

1. Aggregating across a member population 

This implies collecting, normalizing, and storing continuous data from members using different devices, different operating systems, and different data formats. A population of 50,000 members may have heart rate data arriving from thirty different wearable models, each with its own API structure, sampling frequency, and metric definitions. That fragmentation has to be resolved before any population-level analysis is meaningful.

2. Segmenting by risk trajectory rather than diagnosis 

This is where wearable data creates the most differentiated value. Traditional population segmentation uses diagnosis codes to identify high-risk members. Trajectory-based segmentation identifies members whose wearable signals are moving in the direction of a high-risk state, before that state has been clinically confirmed. Pre-patient cohorts, aka members who are not yet diagnosed but whose longitudinal data suggests they are heading toward a diagnosis, represent an intervention window that claims-based models cannot see.

3. Identifying care gaps 

It requires comparing what clinical guidelines recommend for a given member to what their data shows actually happened. A member with hypertension and early heart failure markers should be on a specific medication protocol. If the claims data shows the medication was prescribed but the wearable data shows no change in the expected physiological response, that is a signal worth investigating. The gap between guideline and observed pathway, surfaced at population scale, is one of the highest-value outputs population health analytics can produce.

What AI Does at the Population Layer

Wearable data at population scale generates a volume and dimensionality of signal that cannot be analyzed manually. AI is what makes it actionable.

• Risk stratification at scale applies predictive models trained on combinations of wearable signals, claims, and clinical data to rank a member population by likelihood of a defined outcome: hospitalization, chronic disease progression, readmission, high-cost episode. Stratification models built on wearable features improve predictive sensitivity significantly compared to claims-only models, because they incorporate current behavioral and physiological state rather than historical event data alone.
• Cost trajectory forecasting uses longitudinal member data to project where healthcare spend is heading at the population level. For insurers managing medical loss ratios, a twelve-month cost forecast at the member segment level, informed by current wearable signals, is a different planning tool than a retrospective claims analysis.
• Anomaly detection identifies members who are diverging from their own established baseline in ways that a threshold-based rule would not catch. A member whose resting heart rate has been 58 for eight months and is now at 71 is flagged not because 71 is clinically abnormal in the population, but because it is abnormal for that person. Personalized anomaly detection at population scale is one of the most practically useful applications of AI in this space.
• Care gap detection compares expected care pathways against observed ones across the full member population, identifying systematic gaps in guideline adherence that aggregate to meaningful clinical and financial risk.

What Are the Data Requirements to Make It Work

Population health analytics on wearable data only works if the underlying data infrastructure is built to support it. Several requirements are non-negotiable.

• Normalization across devices: A population where half the members use Apple Watch and half use Garmin will produce HRV data in two different metrics computed by two different algorithms. SDNN and RMSSD are not interchangeable. Any population-level analysis that treats them as equivalent will produce systematically biased results. Normalization has to happen at the infrastructure layer before data enters any analytical pipeline.
• Continuity and completeness: Population models built on sparse or intermittent data are unreliable. Data pipelines need to handle non-wear periods, sync gaps, and device changes without introducing artificial gaps or duplicates that corrupt longitudinal baselines.
• Consent and governance at population scale: Wearable health data is special category data under GDPR Article 9. Processing it for population health analytics requires explicit, purpose-specific consent from each member, and a consent architecture that is queryable at the member level so that new analytical purposes can be managed through proper re-consent workflows rather than retroactive blanket authorization.
• Integration with claims and clinical data: The most powerful population health models combine wearable signals with claims history, diagnostic codes, medication records, and lab results. The infrastructure challenge is joining these datasets at the member level while maintaining data quality, privacy controls, and audit trails across all sources.

How Thryve Powers Population Health Intelligence

Thryve's Population Level Intelligence platform is built on the infrastructure layer described above: normalized wearable data from 500+ connected devices, combined with claims and clinical data, processed through AI models designed for population-scale risk analysis.

The platform surfaces three views of a member population: a population overview with aggregate risk indicators and cost trajectory forecasts, a pre-patient cohort identifying members showing early risk signals before diagnosis, and a chronic patient view that tracks care pathway adherence against clinical guidelines and flags divergence at the member level.

For insurers, the platform identifies care-gap leakage, members whose trajectories are heading toward costly outcomes that earlier intervention could change, and segments where engagement and prevention investment would have the highest return. For clinical teams, it surfaces individual members whose wearable signals are diverging from their established baseline in ways that warrant outreach.

The output is not a dashboard of metrics. It is an actionable view of where risk is accumulating in a population and what can be done about it before the cost is incurred.

If you are building population health capabilities on wearable and claims data and want to understand what the infrastructure and analytics layer looks like in practice, book a demo with Thryve.

‍