Quick Reference

Essential checklists, prompt patterns, and safety guidelines for AI-assisted clinical workflows.

PHI Safety Checklist
Review BEFORE every AI prompt containing clinical information

The "Billboard Test": If this prompt appeared on a billboard, would any patient be identifiable?

Never include in prompts:

  • Patient names (use "Mr. Smith" or "the patient")
  • Medical record numbers (MRN)
  • Dates of birth or exact ages (use "60s" instead of "62")
  • Specific treatment dates
  • Social Security numbers
  • Phone numbers, addresses, email addresses
  • Referring physician names with patient context
  • Any combination that could identify a patient

Safe to include:

  • ✓Fictional patient scenarios
  • ✓Generic clinical details (diagnosis, stage, treatment parameters)
  • ✓Anonymized case presentations
  • ✓General medical questions
  • ✓Template requests without patient specifics
RICE Framework Cheat Sheet
Structure every prompt with these four components
R
Role

Tell the AI who it is

"You are a radiation oncologist..."
"Act as a clinical documentation specialist..."

I
Instruction

What specifically to do

"Write a treatment summary..."
"Summarize this case for tumor board..."

C
Context

Relevant clinical details

"Patient: 65yo male, Stage IIIA NSCLC..."
"Treatment: 60 Gy in 30 fractions via IMRT..."

E
Expectation

Output format, length, tone

"Format as bullet points..."
"Keep under 500 words, professional tone..."

Prompt Patterns for Rad Onc
Copy and adapt these patterns for common tasks

1. Patient Education Letter

You are a radiation oncologist writing to a patient. Write a [reading level]-level explanation of [treatment type] including: what to expect, common side effects, and self-care tips. Tone: warm and reassuring. Length: [word count].

2. Treatment Summary

Draft an end-of-treatment summary for: [clinical details]. Include sections for: Diagnosis, Treatment Delivered (with dose/fractionation), Treatment Tolerance (with CTCAE grades), and Follow-up Plan. Format: structured note suitable for EMR.

3. Dictation Reformatter

Reformat this rough dictation into a structured [visit type] note. Rules: ONLY use information present in my dictation. DO NOT add or infer clinical information. Write "[Not dictated]" for missing sections. Flag errors with [VERIFY:...].

4. Literature Summary (for papers you have)

Summarize this paper's key findings: [paste abstract/text]. Include: study design, population, primary endpoint, key results with numbers, limitations. Note: I will verify all claims before citing.

5. Tumor Board Prep

Extract and organize into tumor board format: ONE-LINER, CLINICAL SUMMARY (table), DISCUSSION POINTS, INFORMATION GAPS. If information is missing, write "[NOT IN SOURCE]" — do not fabricate.
Output Verification Checklist
Check EVERY AI output before clinical use

For Clinical Notes:

  • Doses and fractionation match treatment records
  • Staging matches pathology/imaging
  • Toxicity grades are accurate (CTCAE)
  • Dates are correct
  • Medications are current
  • Follow-up plan matches institutional protocol

For Research/Literature:

  • Citations actually exist (search PubMed)
  • Quoted statistics match source papers
  • No fabricated trial names or authors
  • Dates and journal names are correct
  • Claims are supported by cited evidence
  • Recent guidelines are current (check date)

Remember: AI outputs are DRAFTS, not final products. Apply the same scrutiny you would to a trainee's work.

Common AI Hallucination Patterns
Watch for these red flags in AI outputs

Fabricated Citations

"Smith et al., IJROBP 2023" — May not exist

Action: Always search PubMed before citing

Invented Statistics

"5-year local control of 94.2%" — Precise numbers often fabricated

Action: Verify any specific percentages against source papers

Confident Dose Constraints

"Spinal cord max 45 Gy" — May be outdated or context-specific

Action: Verify against current guidelines (QUANTEC, TG-101)

"Recent" Guidelines

"The latest NCCN recommends..." — AI knowledge has a cutoff date

Action: Check actual publication date of cited guidelines

Plausible Clinical Details

Adding exam findings or assessments not in your dictation

Action: Compare output to your source material line by line