medicine and the machine

What is prompt engineering?

Prompt engineering is an intriguing and increasingly popular concept in the world of generative artificial intelligence. At its core, prompt engineering is the art and science of crafting effective prompts, or instructions, to guide AI models like ChatGPT in generating the most accurate, relevant, and insightful responses. It’s a bit like being a skilled chef; just as a chef combines ingredients to create a delicious meal, a prompt engineer combines words and phrases to extract the best flavors from an AI. This process involves understanding the nuances of language and the capabilities of the AI model, and it’s crucial for applications in healthcare ranging from note-writing, data analysis to image generation. With prompt engineering, we’re not just asking questions; we’re strategically shaping them to unlock the full potential of AI.

Prompt Engineering: Application in Healthcare and Medicine

In the era of advanced artificial intelligence, prompt engineering has emerged as a crucial skill, particularly in healthcare and medicine. This practice involves crafting inputs to elicit the most accurate and relevant outputs from AI models. Let’s explore how prompt engineering can be effectively utilized in healthcare settings, with examples, do’s and don’ts, and practical tips.

Understanding Prompt Engineering

Prompt engineering is essentially the process of designing and refining queries or instructions to interact optimally with AI systems. In medicine, where precision and accuracy are paramount, well-engineered prompts are key to maximizing the potential of an AI chatbot.

Crafting the Right Prompt: A Quick Guide to Becoming a Med Prompt Engineer Using the CO-STAR framework

1. Context (C) Setting the context is the foundation of any interaction with AI. It involves providing the AI with a clear, concise background of what is being asked. This could include the topic, the type of information needed, and any relevant background information.

One common way of easily providing context in medicine is by providing your role in healthcare. You can either indicate in your prompts what your role in healthcare is (doctor, provider, nurse, pharmacist, epidemiologist, students, etc) or you can do it one time in Settings (See our article on this under Setting up ChatGPT).

Example: I’m a nurse treating a patient who just had a heart attack, can you help me explain what a heart attack is to a patient in simple terms?

2. Output (O) Defining the desired output is about being clear on what you expect from the AI. Do you want a detailed report, a summary, or perhaps a list of options? Specifying the format and the level of detail helps the AI deliver precisely what you’re looking for. For example, asking for a concise summary of the notes from previous hospitalizations instead of in-depth review. You can indicate the estimated number of paragraphs or words you would like the chatbot to generate for you in its response.

Example: Summarize the paragraphs below into one 300-word paragraph.

3. Specify (S) Specificity is key in prompt engineering. The more specific your prompt, the more tailored and relevant the AI’s response will be. This involves being clear about the subject matter, the scope of the information, and any particular aspects you’re interested in. In medical queries, specifying whether you’re looking for treatment options, diagnostic criteria, or case studies can make a significant difference in the response.

The chatbot doesn’t know what goes on in your head nor would it know what you need from it, so you have to give clear and specific instructions. Provide additional details or context.

Worse	Better
Rewrite this note:	Rewrite this note in 1 paragraph as a medical transcriber. Be straightforward and professional.
What is the treatment for Enterococcus faecium?	What are intravenous antibiotic options for Enterococcus faecium endocarditis? Give dosing recommendations too. The creatinine clearance is 25.

4. Tone (T) Adjusting the tone of your prompt can yield varied perspectives or responses from the AI. This may involve altering the question’s wording, asking in a different style, or approaching the subject from an unconventional angle.

In medicine, we often communicate with a variety of people, each having different levels of knowledge. We are also expected to adjust our tone depending on the audience. When interacting with our colleagues, a professional tone is expected. Conversely, with our patients, it’s essential to be empathetic and use easy-to-understand language.

Example: Can you explain what hypothyroidism is to a patient in simple terms?

5. Audience (A) The intended audience of the AI’s response can greatly shape the output. Tailoring your prompt based on who will receive the information – be it colleagues, laypersons, or students – ensures that the response is appropriate in complexity, jargon, and depth. In a medical context, the way a question is posed for a fellow physician versus a patient should differ to match their understanding and needs.

6. Response (R) Refining the response or output involves iterative adjustments to your prompts based on the AI’s replies. If the initial response doesn’t fully meet your needs, modifying the prompt with additional details or clarity can help. This process of refinement is crucial in obtaining the precise information you require, particularly in fields like medicine where accuracy is paramount.

You can also indicate the response format to meet your needs. For example, in addition to generating paragraphs, you can indicate the output to be “in bullets” or as “list”. You can also ask the chatbot for a response in a different language.

Others Tips for Effective Prompt Engineering

1. Iterative Refinement: Start with a basic prompt and refine it based on the AI’s responses to hone in on the most effective wording.
2. Use Delimiters to Indicate Different Parts of the Input: Delimiters like line spaces or triple quotation marks (“””insert text here”””) can help the chatbot determine the start and ending of a part of your input.

Example: Translate the text in triple quotes in Spanish. The output should use only simple terms. “””insert text here”””

3. Break it down into Steps: Breaking down prompts into smaller, more manageable steps is a key technique in effective prompt engineering. This is useful for complex tasks or those that require multiple steps.

Example: Step 1: Rewrite the following so that they have the same professional tone. Step 2: Summarize into 1 paragraph.

3. Understand the AI’s Capabilities: Familiarize yourself with the strengths and limitations of the specific AI tool you are using.

4. Collaborate with AI Experts: If available, work with data scientists or AI specialists to craft the most effective prompts.

5. Continuous Learning: Stay updated on the latest developments in AI and machine learning to understand how they can be leveraged in healthcare.

What not to do:

• Avoid Overly Complex Language: This might confuse the AI or lead to irrelevant responses.
• Don’t Underestimate Privacy Concerns: Be mindful of sensitive patient information when formulating prompts.
• Avoid Biased Language: This could lead to biased outputs, which is particularly dangerous in healthcare.
• Avoid Protected Health Information: Never include any identifying information that will violate patient’s privacy. Follow Safe Harbor Method in de-identifying data. Always follow your institutional protocols and regulations when it comes to AI.

Conclusion

Prompt engineering is not just a technical skill but an art that blends understanding of AI capabilities with the nuances of medical knowledge. By mastering this skill, healthcare professionals can unlock the full potential of AI, leading to improved patient outcomes and more efficient healthcare delivery. Remember, the key to successful prompt engineering in medicine is clarity, specificity, and a deep understanding of both the AI system and the healthcare context.

References

OpenAI. (n.d.). Prompt Engineering Guide. Retrieved from https://platform.openai.com/docs/guides/prompt-engineering