🏥

Claims Foundation Model for Medicare

Most US health payers run five to fifteen separate models across admission risk, utilization, fraud, and quality. Each has its own actuarial vendor, its own pipeline, and its own quarterly refresh. This demo collapses three of those into one shared model. Admission risk, next-event forecasting, and claim integrity come back together in under 50 milliseconds, computed from a single read of the beneficiary's claim history, inside the payer's VPC. Adding a new signal like subrogation or Star-rating gap closure is a configuration change. Not a new model build. Shown here on the public CMS DE-SynPUF dataset. Production deployments train on the payer's own claims.

Three signals from one read of the beneficiary. Admission risk over a 90-day window, next-event forecasting, and claim integrity. All three come back together in under 50 milliseconds, inside the payer's VPC.
Reads the trajectory, not just the snapshot. Sees the full sequence of claim events, not aggregated risk scores. Catches dialysis escalation, concentrated-provider billing patterns, and post-discharge readmission windows that flat-feature models cannot represent.
New questions ship in days, not quarters. Subrogation potential, prior-auth steering, Star-rating gap closure. Each becomes a specialized surface on the same shared model. No new pipeline. No new vendor.
Grounded analyst Q&A on the same model. Every prediction comes with a cited summary, ready for the case-manager hand-off or the audit log. Free-text analyst questions are on the roadmap.

The Problem

Payers run five to fifteen separate ML systems across care management, utilization, fraud, and quality. Each has its own vendor, its own pipeline, and its own refresh cadence. Cross-task signals get lost because each model sees only its own features. A beneficiary's dialysis burden that predicts both admission risk and outpatient over-billing is invisible to a model trained on one or the other. Cloud LLMs cannot natively process long structured claim histories.

How LFM Compares

Per-task GBDTs, actuarial lookup tables, or vendor-shipped scoring engines on hand-engineered features. Each new business question means a new feature-engineering cycle, a new model, and a new deployment.

What LFM Unlocks

One model where today there are five to fifteen. Admission, utilization, fraud, quality, and every new signal share the same foundation, the same deployment, and the same audit trail. The model reads raw claim events directly, so the feature-engineering step disappears. VPC-native, so PHI never leaves the perimeter.

Beneficiary claim historyLiquid Claims modelAdmission · Next event · Integrity
Three payer signals from one read of the beneficiary, in under 50 milliseconds.

Beneficiaries

Loading beneficiaries...
Select a beneficiary to see admission risk, next-event distribution, and claim integrity from one model pass.

This demo is fine-tuned on sample data. Results improve with your data.

Build Your Model on Workbench →Talk to Our Team