AI Risk Scoring for Insurers: How Wearable Data Can Support Dynamic Health Models

Traditional health risk scoring is a snapshot taken at the wrong moment. It relies on claims history, self-reported questionnaires, and periodic check-ups, data that is backwards-looking, infrequent, and structurally incomplete. A member who was healthy at their last assessment three years ago may be sedentary, sleep-deprived, and showing early cardiovascular signals today. The model does not know, because the model has not been updated.

Wearables change the underlying data availability entirely. The physiological signals that matter most for health risk, like resting heart rate, heart rate variability, sleep quality, activity levels, are now being generated continuously by devices that millions of people already wear every day. The question for insurers is no longer whether this data exists. It is how to turn it into risk models that are actually predictive, operationally sound, and compliant with the privacy requirements that govern health data.

How Traditional Risk Models Fall Short

Most health insurance risk models are built on claims data: diagnoses, procedures, prescriptions, and hospitalizations. These records are rich in clinical detail but structurally delayed. A claim reflects something that has already happened. By the time deteriorating health shows up in the claims record, a preventable event has often already occurred.

Questionnaire-based risk assessments add some behavioral context, but they are self-reported, completed infrequently, and subject to recall bias and social desirability effects. A member asked once a year whether they exercise regularly is not the same as a member whose activity levels are measured daily.

The combined result is a risk model that is most reliable for members who have already generated significant claims history, precisely the members whose risk is already well understood. For members who are currently healthy but deteriorating, or for newly onboarded members with limited data history, traditional models offer limited predictive power exactly where it is most valuable.

How Wearable Data Improves Risk Assessment 

Wearables provide a continuous stream of physiological and behavioral data that traditional models cannot access. The signals most relevant to health risk are:

• Resting heart rate: a well-validated cardiovascular risk indicator, with elevated resting heart rate associated with increased mortality and cardiovascular event risk
• Heart rate variability (HRV): a marker of autonomic nervous system function, with declining HRV associated with stress, poor recovery, and early-stage cardiovascular disease
• Sleep duration and quality: chronic sleep disruption is linked to metabolic dysfunction, immune suppression, and elevated risk for diabetes, hypertension, and cardiovascular disease
• Physical activity and sedentary behavior: step counts, active minutes, and time spent sedentary provide a continuous measure of lifestyle that questionnaires approximate only coarsely
• Blood oxygen saturation: relevant for respiratory risk monitoring, particularly in populations with known or suspected pulmonary conditions
• Skin temperature trends: an emerging signal for early illness detection and menstrual cycle tracking, with broader physiological relevance under investigation

None of these signals are new to clinical medicine. What is new is the ability to observe them continuously, at population scale, outside of a clinical setting.

How AI Turns Biometric Signals Into Risk Scores

From Raw Data to Predictive Features

Raw wearable data signals, like a sequence of heart rate readings, a nightly sleep record, are not directly usable as a model input. The predictive signal lives in patterns over time: a resting heart rate that has been gradually increasing over three months, HRV that is declining relative to a personal baseline, sleep efficiency that drops consistently below a threshold for extended periods.

Feature engineering for wearable data typically involves:

• Baseline establishment: computing individual baselines over a sufficient observation window before deviations become meaningful
• Trend detection: identifying directional changes over days, weeks, and months rather than point-in-time values
• Variability measures: quantifying how stable or unstable a given signal is over time, which often carries as much information as the signal level itself
• Cross-metric patterns: combinations of signals (declining HRV alongside rising resting heart rate and worsening sleep) that are more predictive than any single metric alone

Model Approaches in Practice

The most effective approaches combine wearable-derived features with traditional data sources rather than replacing one with the other.

• Gradient boosting models (XGBoost, LightGBM) handle structured tabular data well and are widely used for integrating wearable-derived features with demographic and claims data
• Time-series models capture the temporal structure of biometric data more naturally, identifying patterns that tabular approaches may miss
• Hybrid architectures use deep learning components for raw sequence data and gradient boosting for the final risk prediction, combining the strengths of both

The output is a risk score that reflects both historical claims context and current physiological state and that updates as new data arrives.

What Dynamic Actually Means

A truly dynamic risk model does not produce a score at enrollment and revisit it annually. It produces a score that is recalculated as the underlying data changes, reflecting behavioral improvements, detecting early deterioration, and flagging members whose risk profile has shifted significantly since their last formal assessment.

This is where wearable data creates a structural advantage over traditional approaches. Claims data updates when a claim is filed. Wearable data updates every day.

Where Wearable-Based Risk Scoring Creates Value for Insurers

Underwriting

Earlier and more accurate risk stratification at the point of underwriting reduces adverse selection and improves pricing accuracy. For members who consent to wearable data sharing, physiological signals provide a more current picture of health status than a medical history review or a one-time health screening. That’s why, at Thryve, we offer the Swift Intel Risk Model that allows insurers to maximize the benefits. 

Prevention and Early Intervention

The highest-value application of continuous risk scoring is not identifying who is already high-risk, it is identifying who is becoming high-risk before a clinical event occurs. A member whose HRV has been declining for eight weeks alongside worsening sleep and reduced activity is generating a signal that a claims-based model will not see for months, if ever. Identifying that member now, and enabling a targeted intervention, is where wearable-based risk models create measurable clinical and financial value. Thryve’s Prevention Pathway Guide allows you to utilize the full potential of your preventive solutions! 

Engagement and Loyalty

Incentive programs tied to behavioral data create a feedback loop that benefits both the member and the insurer. Members who engage with wellness programs generate more data, enabling better models. Better models enable better-targeted interventions. And insurers who offer meaningful, data-driven engagement programs retain members more effectively than those who do not. To be ahead of the curve, check out our Member Connection Engine!

The Challenges That Need to Be Solved

Data Quality and Device Fragmentation

Wearable data is only as useful as it is reliable. HRV measured by a WHOOP uses a different methodology than HRV measured by an Apple Watch or a Garmin. Sleep staging algorithms are proprietary across platforms. Without normalization, aggregating data from a member population using mixed devices produces noise rather than signal.

Normalization has to happen at the infrastructure level. Attempting to handle it per device in application code produces brittle, unmaintainable logic that breaks with every API update.

Privacy and Regulatory Compliance

Health data is special category data under GDPR Article 9, subject to stricter processing requirements than ordinary personal data. For insurers operating in European markets, wearable-based risk programs require:

• Explicit consent from members for health data collection and processing
• Clear purpose limitation: data collected for wellness purposes cannot be repurposed for underwriting without separate consent
• Data minimization: collecting only the signals necessary for the stated purpose
• EU data residency: ensuring health data is processed and stored within the EEA

Regulatory requirements are not a reason to avoid wearable data programs. They are a reason to build them on infrastructure that handles compliance by design rather than retrofitting it afterward.

Adverse Selection in Opt-In Programs

Wearable data programs are typically opt-in. The members most likely to opt in are often those who are already health-conscious and physically active,  precisely the members whose risk is lowest. Models trained predominantly on this subpopulation may not generalize reliably to the broader member base.

Addressing this requires careful program design: incentive structures that attract broader participation, model validation across demographic subgroups, and explicit tracking of participation bias in model performance metrics.

What the Infrastructure Has to Support

Building a wearable-based risk scoring capability requires infrastructure that most insurance organizations are not starting with:

• Normalized data across devices: so that a member population using mixed devices produces comparable signals
• Compliant data collection: consent management, purpose limitation, and data residency built in from the start
• Reliable, scalable data delivery: webhook-based delivery and API access that performs consistently at population scale
• Clean interfaces into model pipelines: structured, normalized data that feeds directly into feature engineering and model training workflows without requiring device-specific preprocessing

How Thryve Supports Insurer Risk Models

At Thryve, we build the infrastructure that makes wearable data operationally usable for insurers and digital health platforms. Our wearable API connects to 500+ wearables and health data sources and returns normalized, structured data through a single integration with GDPR-compliant data handling built in at the infrastructure level.

With our platform, you get:

• Normalized biometric data across all connected devices, with consistent metric definitions for HRV, sleep, activity, heart rate, and more
• GDPR-compliant infrastructure with consent management, EU data residency, and data subject rights supported out of the box
• Health risk assessment capabilities that translate normalized wearable data into actionable health insights, ready to feed into risk scoring models
• Scalable data delivery via webhooks and API, designed for population-scale programs rather than individual user tracking

If your organization is building or evaluating wearable-based risk programs, book a demo with Thryve to see how the infrastructure layer works in practice.