Apply rules
A synthetic 1,000-patient panel is screened against USPSTF and HEDIS rules. Each open gap is annotated with which guideline triggered it and what evidence rests behind that guideline.
01 · What it does
Closing the gap between care recommended and care delivered.
Primary-care teams under HEDIS or value-based-care contracts don't fail because they lack guidelines. They fail because they have too many patients with open gaps and a finite week. The Care Gap Engine treats outreach as a scheduling problem: which patient, with which gap, gets a message this week, and what does that message say.
The engine reads from a synthetic patient panel, applies USPSTF preventive-care rules and HEDIS quality measures, scores each gap on a three-component priority surface, and drafts outreach for the top-ranked patients. The synthetic panel keeps the demo runnable without PHI; the same shape of pipeline plugs into a real EHR with a thin adapter.
02 · How it works
Three stages.
Score and rank
Three components combine into a single priority score: clinical urgency, response likelihood (based on prior outreach behavior), and equity priority. Weights are tunable per contract.
Draft outreach
Claude drafts personalized outreach for each top-ranked patient. Prompt caching keeps per-message cost roughly flat across hundreds of drafts.
03 · Behind the demo
Methodology.
Rules are split between USPSTF preventive-care recommendations (cancer screening, immunizations, counseling) and HEDIS quality measures (control of chronic conditions, follow-up after hospitalization). Each triggered gap carries a citation back to the source rule, so a clinician reading the prioritized list can sanity- check why a patient surfaced.
The priority score combines clinical_urgency
(how time-sensitive the gap is), response_likelihood
(a per-patient estimate from prior outreach history), and
equity_priority (a weight applied to patients
from underserved groups so outreach budget doesn't
concentrate on the easiest-to-reach). Weights are tunable
per contract, defaults are documented in the repo.
Outreach drafting uses Anthropic prompt caching: the system prompt and patient-context preamble are cached, and only the per-patient delta is billed at the full input rate. Across hundreds of drafts the per-message cost stays roughly flat, which is what makes the engine tractable as panels scale.