Designing for Clinical Safety
Support clinical safety in AI-assisted note drafting by linking every clinical claim to its source, requiring clinician approval of the exact draft version, and conducting systematic evaluations to identify omissions and inaccuracies. Establish a structured workflow that integrates audio capture, transcription, draft generation, and clinician review. Verify that all drafted claims accurately reflect the source material and prevent unsupported or misinterpreted information from entering medical records. Implement a review interface that enables clinicians to validate, correct, and approve documentation before finalization. Monitor errors, omissions, and reviewer performance after deployment, and trigger immediate clinical review when significant discrepancies are detected. Maintain the accuracy, integrity, and safety of clinical documentation through continuous evaluation and improvement.
How do you know a clinical draft is safe to file? Not because it cites a source: a citation can point straight at a sentence that says the opposite. Safety comes from three design choices working together. Each claim in the draft is tied to the source text, a clinician approves one exact version before it can be filed, and evals look specifically for omissions and invented findings. You’ll build the first two as a small runnable program, then design the evals and the monitoring that would follow a launch.
How does the drafting pipeline start?
A production pipeline for Lakeside would capture audio, transcribe it, select the relevant transcript text, request a draft, and show it for review. This design starts after transcription, with a short test transcript. Transcription still needs its own approval and quality checks, plus a way to trace each transcript line back to the audio.
Each transcript turn gets a stable ID, a speaker, an order, and the source version S1:
T1 (patient): “I have not had a fever.”
T2 (patient): “The nausea started yesterday.”
T3 (clinician): “It may be related to the new medicine, but I am not sure.”
The encounter contains no allergy history. That’s the gap that produced “No known medication allergies” in the earlier demonstration, and missing history supports neither a positive nor a negative allergy finding.
How are claims tied to turns?
Represent the draft as a list of claims, each with its text and the turn IDs that support it. Before reading the examples, decide which claim the encounter supports.
A faithful draft says “The patient reports no fever” citing T1, “The patient reports that nausea started yesterday” citing T2, and “The clinician said the nausea may be related to the new medicine, but was unsure” citing T3. Each claim keeps something that matters clinically: T1 keeps the negation, T2 keeps the timing, and T3 keeps both the speaker and the uncertainty.
Now look at two fluent alternatives. “Patient has a fever” [T1] has a citation, but T1 says the opposite. “The patient confirms the new medicine caused nausea” [T3] turns a clinician’s uncertain guess into a patient’s certain statement. The link tells the reviewer where to look; the reviewer still has to compare meanings.
The fever claim shows why a citation alone proves nothing. The allergy claim makes a different mistake: it treats missing history as a negative finding. Fix a claim so it keeps its source’s meaning, or remove it when there’s no source.