A twelve-month behavioral risk score, written in claims, cadence, and continuous glucose.

Why a behavioral risk score, in twelve months

The decision a benefits leader can actually act on is rarely "is this person sick now." It is "who in my population is on a trajectory toward a high-cost condition in the next budget cycle." BR2H-Score was designed for that question — a forward-looking, twelve-month horizon that surfaces the 4.7% of a workforce most likely to convert to a high-cost event before a diagnostic code is ever written. It is a population-health instrument, not a clinical diagnostic.

Three signal families, fused at inference

The model ingests, in a single HIPAA-compliant lake, three signal families that previously lived in three different systems: passive wearable streams (continuous glucose, HRV, step-cadence, sleep architecture), adjudicated medical and pharmacy claims, and structured EHR fields via FHIR R4. At inference these are not concatenated — they are fused through a Bayesian residual architecture that respects each stream's missingness pattern and time resolution, so a member who only consents to wearable sharing still receives a calibrated score.

Single-tenant by default, with sub-150ms inference on PHI

Every BR2H deployment is single-tenant. PHI never crosses tenant boundaries, model weights remain inside the customer's HITRUST CSF v11 boundary, and inference latency on protected data is benchmarked under 150 milliseconds. SOC 2 Type II, ISO 27001, and HIPAA compliance are audited annually by Coalfire; uptime across 2023 production deployments was 99.982%, verified by independent monitoring.

Co-developed with Stanford, validated against 2.4M lives

The model was co-developed with the Stanford Prevention Research Center and has produced 47 peer-reviewed publications since 2021. It is validated across the full 2.4-million-member BR2H covered base, and its 30-day readmission prediction reaches 91.3% AUC — a figure that has been independently reproduced, not self-reported.