Carta Healthcare's Claude-based Lighthouse cut clinical registry abstraction time by up to 66%

What the agent does

Hospitals report data to clinical registries (for example, cardiac surgery outcomes). Trained staff called abstractors read charts to answer each registry question. A routine case can take an hour; a complex one five or six hours.

Carta's Lighthouse platform uses Claude in two steps. Smaller models pull relevant facts from notes, labs and medication records. A larger model weighs the evidence and drafts an answer with a reason and direct citations. Abstractors see the evidence and rationale and make the final call.

Why this works

Carta says the key is context, not clever wording. For each question, its pipeline gives Claude the right documents, time window and order. For a pre-procedure glucose value, the prompt includes the procedure start time, so Claude looks only for the last value before it. When abstractors spot a mistake, their explanation becomes a prompt fix, often the same day.

What to copy

• Give the model a precise time window and source list for each question.
• Show reviewers the evidence and the reasoning, not just the answer.
• Turn reviewer corrections into prompt changes quickly.

Starter prompt

Written by us, not by Carta Healthcare.

Registry question: {question}
Time window: from {start} to {end}
Documents (with dates): {documents}

Answer using only documents inside the time window.
Give: answer, the exact quote, the document name and date, and one sentence of reasoning.
If two documents disagree, show both and say which you chose and why.
If nothing answers the question, say "No evidence found".