One risk score anyone can read the reason for

The problem

The member roster had a risk column that was really a placeholder with no data behind it. Its first replacement disagreed with the score on the patient page, where one old critical flag kept a patient high indefinitely. Two numbers for the same patient is worse than none.

What was built

Each patient gets one 0 to 100 risk score. It blends three things: recent surveillance flags (weighted by severity and fading over time), the patient's spending trend, and their value-based-care deficit. The breakdown sits in the patient header, so anyone can see what moved the number.

The surveillance flags come from rules checked against every claims upload: high-cost claims, unusual monthly totals, statistical outliers. Thresholds are editable, previewed before they are published to everyone, and older flags are recomputed when a rule changes.

What was hard

Choosing the weights. The value-based balance ended up carrying half the score (30% flags, 20% spending trend, 50% balance), so a deep and worsening deficit on its own lifts a patient to about the middle of the scale.

A trap in the data: the newest month in every claims file is incomplete, because medical claims keep arriving for one to three months. It cannot be compared like a full month, and anyone reading surveillance numbers has to know that.

Result

One score, one explanation, the same on every screen. The outcome that would prove it useful, the share of the 50 highest-risk patients with an open case, is defined but not yet measured, so no result figure is claimed.

Where the numbers come from

Build notes from July 2026. The weights are design decisions, recorded as such.