A healthcare network says physicians lose about two hours a day to clinical documentation. How would you scope a project to reduce that?
Answer it out loud first, then compare. That's what makes it stick.
First I'd find where the two hours actually go: typing notes, order entry, inbox, coding. I'd shadow a few physicians across specialties, because needs differ. Then I'd ask about the EHR and its integration options, privacy rules, who owns note quality, and how they'd measure success. A first deliverable might be a draft note for one specialty in one clinic, reviewed and signed by the physician, tracked by minutes saved and edit rate.
- Break down the two hours: where does time go?
- Shadow different physician types
- Integration, privacy and ownership constraints
- Success metric: minutes saved, edit rate, safety
- One specialty, one clinic, human sign-off
This is one of the prompts a published guide lists for the decomposition round. It is a classic because 'documentation time' is a symptom with several causes.
Step 1: Decompose the two hours. Ask for a time study or do one: 5 physicians shadowed for half a day. Typical buckets are writing the visit note, reviewing records before the visit, order entry and referral forms, inbox messages, and after-hours charting. Each has a different fix, and only some involve language models.
Step 2: Segment users. A primary-care doctor, a surgeon and an emergency physician document differently. Start with the group that has the biggest pain and the most standard note format.
Step 3: Constraints to ask about.
- EHR and integration: can you read the record and write a draft back, or is it copy and paste at first?
- Privacy and compliance: is patient data allowed to leave their environment, who must approve vendors, what audit trail is needed?
- Liability: the physician must remain the author who signs. Design for review, never silent filing.
- Ownership: who owns note quality (compliance, a CMIO, a clinical champion)?
Step 4: Define success.
- Primary: minutes of documentation per visit, measured before and after.
- Quality guardrails: edit rate of drafts, factual errors found in audit (for example a missed allergy), and physician satisfaction.
- Do not claim the full two hours; a realistic first target is a fraction of it, and you should say so.
Step 5: First deliverable. One specialty, one clinic, 5-10 volunteer physicians, 4 weeks. The tool drafts the note from the visit; the physician edits and signs. Log edits, because the edits are your best training signal and your eval data.
Step 6: Risks with fixes.
- Hallucinated clinical facts: require source links to the transcript or record and a visible 'unverified' marker.
- Physician fatigue from reviewing bad drafts: gate rollout on edit rate.
- Privacy review delay: begin with a sandbox on synthetic or de-identified data.
Say clearly what you would not do first: automate coding or orders, which have higher risk and more rules.
- Physicians are split: half want it, half refuse. What do you do?
- How would you measure error risk before the pilot?
- The CMIO wants to ship to all clinics in a month. How do you respond?
- The four interview rounds (Educative) (prep, unknown) Lists the healthcare documentation-time prompt as an example decomposition problem with a 60-minute format.
- Decomposition interview guide (prep, unknown) Rubric: clarify before solving, make assumptions explicit, reason about trade-offs, keep the end user in view.