Clinical Decision Support

A suggestion at the bedside is not a recommendation until a clinician makes it one. This starter wraps an assistant that helps with that moment. It masks patient identifiers, holds the answer to a numeric quality floor, and leaves a review point your team must activate. It makes no claim about clinical validity or device clearance.

A clinical informatics lead reaches for this starter when an assistant sits beside an order workflow. The risk is not only exposure of patient data. The larger risk is a confident answer that a clinician accepts without checking the underlying record. Controls here aim at both problems in one pass.

Screening for prompt injection runs first, so a crafted instruction cannot disable the identity check behind it. The identity check then refuses any call that arrives without verified user and organization claims. Health record heuristics run next and add clinical detection on top of the shared redaction pipeline. Masking follows for patient numbers, member record numbers and insurance or policy numbers. All of this finishes before the model receives the prompt.

Citation checking measures the answer against supplied context and flags weak grounding rather than refusing it. Output scoring is stricter. It holds a combined score of 0.88, with relevance at 0.78 and accuracy at 0.80. An answer under 120 characters or a single sentence fails the same gate, and the default result is a block.

Two limits matter before rollout. The review point ships without an active escalation rule, so a team must turn it on to route an answer to a person. The audit step marks an evaluated request and never blocks one, and it does not by itself store or retain anything. History capture keeps metadata only, which suits a site that must not hold raw clinical text. Pick the Medical Records Summarization starter when the task is chart abstraction rather than a decision at the bedside.