AI Fundamentals for Radiation Oncology
Understand how large language models work, their capabilities and limitations, and key considerations for clinical use.
- Explain how LLMs generate text at a high level
- Identify appropriate vs inappropriate use cases
- Understand privacy and HIPAA implications
- Recognize common AI limitations and pitfalls
Large Language Models (LLMs) like GPT-4, Claude, and Gemini are AI systems trained on vast amounts of text data. They learn patterns in language that allow them to generate human-like text responses.
Key Concepts:
- Training: LLMs are trained on billions of words from books, websites, and documents. They learn statistical patterns about how words and concepts relate to each other.
- Prediction: At their core, LLMs predict the most likely next word (or token) given the preceding context. This simple mechanism, scaled up massively, produces remarkably coherent text.
- No True Understanding: Despite impressive outputs, LLMs don't "understand" in the human sense. They pattern-match based on training data, which is why they can produce confident-sounding but incorrect information.
Analogy for Clinicians:
Think of an LLM like an extremely well-read medical student who has memorized vast amounts of medical literature but lacks clinical experience. They can discuss concepts fluently and even make reasonable-sounding suggestions, but they haven't actually seen patients or verified their knowledge against real outcomes. You wouldn't let them practice unsupervised, and the same principle applies to AI outputs.
LLMs excel at certain tasks that are particularly relevant to radiation oncology practice:
Strong Capabilities:
- Text Summarization: Condensing long documents, papers, or notes into concise summaries.
- Writing Assistance: Drafting clinical notes, letters, and documentation with appropriate medical terminology.
- Translation & Simplification: Converting medical jargon into patient-friendly language.
- Information Synthesis: Combining information from multiple sources into coherent narratives.
- Template Adaptation: Taking a template and filling it with specific patient or scenario details.
- Brainstorming: Generating ideas, differential diagnoses (for educational purposes), or discussion points.
Example: Treatment Summary Draft
A radiation oncologist might use an LLM to draft an initial treatment summary:
Prompt: "Draft a brief treatment summary for a patient who completed 60 Gy in 30 fractions of IMRT to the prostate with concurrent ADT. The treatment was well-tolerated with grade 1 urinary frequency."
The LLM can produce a well-structured draft that the physician then reviews and modifies with patient-specific details.
Understanding LLM limitations is essential for safe clinical use:
Major Limitations:
- Hallucinations: LLMs can generate plausible-sounding but entirely false information—fake citations, incorrect drug doses, or imagined clinical trials.
- Knowledge Cutoff: LLMs have training cutoff dates and don't know about recent guidelines, drugs, or research.
- No Verification: LLMs cannot verify facts against current databases or literature. They generate text based on patterns, not truth.
- Context Window Limits: LLMs can only "remember" a limited amount of text in a conversation, potentially missing earlier context.
- Inconsistency: The same prompt may generate different outputs, and small prompt changes can dramatically alter results.
- Bias: Training data biases (demographic, geographic, temporal) are reflected in outputs.
Red Flags to Watch For:
- Specific citations (authors, journals, years) - these are often fabricated
- Precise numerical values (doses, survival statistics) - verify independently
- Claims about "recent" or "latest" guidelines - check currency
- Confident statements about rare conditions - pattern matching may fail
- Any output you would use for patient care - always verify
Never enter Protected Health Information (PHI) into public AI systems.
PHI includes: names, dates, medical record numbers, addresses, phone numbers, email addresses, Social Security numbers, and any other information that could identify a patient.
Safe Practices:
- Use fictional patient scenarios for learning
- Fully de-identify any real cases before analysis
- Check your institution's AI use policies
- Consider institution-approved AI tools with BAAs (Business Associate Agreements)
- When in doubt, don't include patient information
What You CAN Safely Ask:
- General medical questions
- Help with template structures
- Explanations of concepts
- Editing anonymized text
- Writing exercises with fictional cases
Remember: Once information is sent to an AI system, you have limited control over how it's stored or used. Treat AI prompts like unencrypted emails—assume they could be seen by others.
For each scenario, consider whether it would be appropriate to use a public LLM (like ChatGPT or Claude):
- Drafting a patient letter explaining what to expect during external beam radiation therapy for breast cancer. You'll review and edit the draft before sending.
- Looking up the recommended dose constraints for spinal cord in SBRT from AAPM TG-101.
- Summarizing a specific patient's treatment history by pasting their clinical notes into the AI.
- Generating ideas for a departmental quality improvement project on reducing treatment delays.
- Getting the "current standard of care" for newly diagnosed glioblastoma.
Answers:
- ✅ Appropriate - Generic patient education content with physician review is a good use case.
- ⚠️ Use with caution - LLMs may have outdated or incorrect dose constraints. Always verify against the actual guideline document.
- ❌ Not appropriate - Contains PHI. Never paste patient records into public AI systems.
- ✅ Appropriate - Brainstorming for QI projects is a good use case for idea generation.
- ⚠️ Use with caution - Treatment standards evolve. LLM knowledge may be outdated. Always verify with current NCCN guidelines or institutional protocols.
Test your understanding of LLM fundamentals:
Q1: When an LLM provides a specific citation (e.g., "Smith et al., IJROBP 2023"), you should:
- A) Trust it because LLMs are trained on medical literature
- B) Verify the citation exists before using it
- C) Assume it's approximately correct
- D) Only verify if it seems unusual
Q2: Which task is MOST appropriate for LLM assistance?
- A) Determining the appropriate radiation dose for a specific patient
- B) Drafting a template for weekly on-treatment visit notes
- C) Looking up a patient's allergy history
- D) Calculating a patient's GFR from lab values
Q3: The primary reason LLMs can "hallucinate" false information is:
- A) They are designed to be creative
- B) They predict likely text patterns, not factual truth
- C) Their training data contains errors
- D) They try to please the user
Answers:
Q1: B - Always verify citations. LLMs frequently generate plausible-sounding but non-existent references.
Q2: B - Template drafting is ideal—it's generic, doesn't require patient-specific accuracy, and will be reviewed before use.
Q3: B - LLMs are fundamentally prediction engines. They generate statistically likely text, which may or may not be factually accurate.
Key Takeaways:
- LLMs are powerful text tools, not knowledge oracles. They generate plausible text based on patterns, not verified facts.
- Excellent for drafting and editing, poor for fact-checking or clinical decision-making.
- Always verify any specific claims, citations, doses, or guidelines against authoritative sources.
- Never enter PHI into public AI systems. Use fictional scenarios or fully de-identified information.
- Think of AI as a first draft, not a final answer. Your clinical judgment and verification are essential.
Next Steps:
Now that you understand the fundamentals, continue to the Clinical Note Writing module to learn practical techniques for using AI in documentation workflows.