VyaHealth Insights

The science behind a single number for your health

We built the VyaHealth Score™ on data from over 18,000 U.S. adults and stress-tested it against measured HbA1c, blood pressure, and lipid panels. Here's what we found — in plain English, with the numbers behind it.

0.844
Internal accuracy
BRFSS 2024 held-out test set
0.937
Real-world accuracy
NHANES 2021–2023, no retraining
0.821
Predicts HbA1c diabetes
From a self-report intake
0.925+
Equity floor
All race/ethnic subgroups in NHANES
What it is

A 0–100 read on where your preventive health stands today

Most digital health tools live in silos: one app for steps, another for blood pressure, a separate calculator for cardiovascular risk, another for diabetes. None of them tell you how you're doing overall.

The VyaHealth Score™ is a single index — higher is better — that combines modifiable lifestyle, current medical burden, preventive-care engagement, and (in Phase 2) family history into one interpretable number. It's designed to be understandable to you and credible to your clinician.

It is not a diagnosis. The score is a prevention-intelligence layer — it surfaces drivers, points to next actions, and helps you and your care team prioritize. Clinical decisions still belong in the exam room.

What feeds the score

Four domains, weighted by what actually moves outcomes

Domain weights are anchored in evidence. The CDC estimates that ~40% of deaths can be traced to modifiable behaviors.

Lifestyle & Behavioral Health

20–30%

Smoking, BMI, physical activity, sleep, alcohol. The most modifiable inputs — the things you can actually change.

Personal Medical Profile

15–20%

Chronic conditions, days of poor physical and mental health, functional limitations. Where your current burden sits.

Prevention Readiness

20–25%

Checkup recency, vaccinations, age-appropriate screenings. How engaged you are with preventive care.

Family & Hereditary Risk

20–25%

Family history of major chronic diseases and hereditary risk flags. Added in Phase 2 through the MyVya360 intake.

How we built it

Trained on 10,000 U.S. adults. Held out a third. Then tried it on a totally different dataset.

The score was developed on the CDC's 2024 Behavioral Risk Factor Surveillance System (BRFSS) — 457,670 records, from which we drew a survey-weighted analytic sample of 10,000.

  • 70/30 split. A survey-weighted logistic regression learned from 7,000 respondents. The other 3,000 were held out and never touched until final scoring.
  • External validation. The same trained model — with no retraining or tuning — was then applied to NHANES 2021–2023, an entirely separate national survey with measured labs.
  • Pre-registered thresholds. An AUC drop > 0.05 between train and test would have flagged overfitting. The actual difference was −0.003.
Figure 1 — BRFSS 2024 test set

People in the highest risk band are nearly 10× more likely to report fair or poor health

As the VyaHealth Score drops, the share of people reporting fair or poor general health rises sharply — a clean, monotonic gradient across all five risk bands.

Low · score 80–1007.2%
Moderate-Low · score 65–7918.4%
Moderate · score 50–6434.6%
Moderate-High · score 35–4952.1%
High · score 0–3468.4%

Source: BRFSS 2024 held-out test set (n = 3,000). Outcome: RFHLTH (self-reported fair or poor general health).

Figure 2 — NHANES 2021–2023 MEC exam

The score predicts measured lab outcomes, not just how you feel

Applied without any retraining to NHANES, the VyaHealth Score discriminated diabetes from blood-drawn HbA1c at AUC 0.821 — stronger than many single-disease calculators designed only for that purpose.

HbA1c ≥ 6.5% (diabetes)AUC 0.821
Measured obesity (BMI ≥ 30)AUC 0.762
Composite cardiometabolic riskAUC 0.660
Elevated LDL cholesterolAUC 0.612
Hypertension (measured BP)AUC 0.571

AUC of 0.5 is chance; 1.0 is perfect. Source: NHANES Aug 2021–2023 MEC exam (n = 6,337). Composite Cardiometabolic Risk Flag (CCRF) is a 7-criterion clinical composite.

The hard test

Can a self-report score predict what a blood test would say?

This is the part most "wellness scores" never publish. We checked whether the VyaHealth Score — built without any lab data — could still pick out people who had clinically meaningful lab abnormalities in NHANES.

For undiagnosed diabetes defined by measured HbA1c ≥ 6.5%, discrimination was AUC 0.821 — competitive with purpose-built diabetes risk calculators. The composite cardiometabolic risk flag landed at AUC 0.660, with prevalence rising from 71.6% in the lowest VyaHealth band to 93.6% in the highest.

Equity by design

When a model fails a group, name the missing data

Most health-AI fairness reports stop at "there's a gap." We went further: when the Hispanic subgroup AUC dropped to 0.688 in the first model, we asked why. The answer was in the inputs — access-to-care variables like insurance coverage, usual source of care, and cost barriers were not included.

Adding those access-to-care signals lifted Hispanic AUC to 0.925 and brought every evaluated subgroup above 0.92. Those variables are now mandatory in the MyVya360 intake.

Figure 3 — Equity across race & ethnicity

Adding access-to-care signals closes the equity gap

The first model did not include access-to-care variables — insurance, usual care, or cost barriers. When those signals were added, the Hispanic subgroup AUC jumped from 0.688 to 0.925 — closing the largest fairness gap in the model.

White, non-Hispanic
BRFSS
0.842
+ NHANES
0.938
Black, non-Hispanic
BRFSS
0.811
+ NHANES
0.929
Hispanic
BRFSS
0.688
+ NHANES
0.925
Other / multiracial
BRFSS
0.821
+ NHANES
0.930

AUC by subgroup. The validated model includes access-to-care variables (insurance, usual source of care, cost barriers) that close the equity gap.

Methods, in brief

For the analytically minded

Model

Survey-weighted logistic regression, 41 predictors across three BRFSS-scoreable domains. Newton–Raphson optimization, Hessian-based standard errors. VIF < 2.5 across all predictors.

Outcome

RFHLTH — a binary indicator of self-reported fair or poor health, derived identically in BRFSS and NHANES. Validated in 20+ years of prospective studies as a robust predictor of mortality and decline.

Calibration

Brier Skill Score 0.303 (≈ 30% improvement over a prevalence-only baseline). Modest over-prediction at lowest deciles targeted for Platt scaling in Phase 2.

Events per predictor

33.7 — well above the conventional minimum of 10, supporting stable coefficients without regularization.

Data sources: CDC Behavioral Risk Factor Surveillance System (BRFSS) 2024, n = 457,670; National Health and Nutrition Examination Survey (NHANES) August 2021–2023, n = 8,153 interview / 6,337 MEC exam. The full white paper — VyaHealth Score™: A Preventive Health Index for Personalized Preventive Intelligence (VyaHealth Insights Series, Report No. 1) — is available on request.

See your own VyaHealth Score™

Under 60 seconds. No lab tests. No insurance hassles. The same model described on this page — applied to you.