Medical Records Summarization

Someone has to read forty pages of chart before a handoff, a prior authorization or a referral letter. A model can compress that in seconds. The risk is that the compression stays silent about what it dropped. This starter puts identifier masking on the way in and a scored gate on the way out.

Chart abstraction fails in a specific way. A summary reads well, omits one material finding, and nobody notices because the source chart is long. Length and fluency are not accuracy. The response side of this starter is built around that gap.

Two response checks do most of the work. Source checking pulls citations out of the summary and measures them against the context provided, then flags weak grounding without refusing the text. Quality scoring is the gate that refuses. That gate asks for 0.88 combined, 0.78 for relevance and 0.80 for accuracy. It also rejects anything under 120 characters or shorter than two sentences.

The request path is shorter to describe. Injection screening protects the checks behind it, identity enforcement rejects an unauthenticated caller, and masking hides patient, member and insurance numbers. A health data stage contributes chart oriented heuristics to that same masking step. A narrative note that names a person without any number is not covered.

Know what stays open. Human review is present as a configuration point with no rule inside it, so no summary routes to a reviewer automatically. Metadata only history keeps chart text out of the captured record. The audit marker never blocks and never becomes a storage system on its own. Reach for the Clinical Decision Support starter instead when an assistant informs an order at the point of care. Both starters carry the same control set, so pick by workflow and not by strength.