Designing Prompts for Generative Artificial Intelligence in Clinical Oncology Contexts

Abstract

AI in Precision OncologyAhead of Print Prompt AssistanceFree AccessDesigning Prompts for Generative Artificial Intelligence in Clinical Oncology ContextsDouglas B. Flora and Nikhil G. ThakerDouglas B. FloraEditor-in-Chief, AI in Precision Oncology.Search for more papers by this author and Nikhil G. Thaker*Address correspondence to: Nikhil Thaker, MD, MHA, MBA, Department of Radiation Oncology, Capital Health Radiation Oncology, One Capital Way, Pennington, NJ 08534, USA, E-mail Address: [email protected]Department of Radiation Oncology, Capital Health Radiation Oncology, One Capital Way, Pennington, New Jersey, USA.Search for more papers by this authorPublished Online:17 Oct 2023https://doi.org/10.1089/aipo.2023.0004AboutSectionsPDF/EPUB Permissions & CitationsDownload CitationsTrack CitationsAdd to favorites Back To Publication ShareShare onFacebookXLinked InRedditEmail As we stand at the forefront of a paradigm shift in oncology, the intertwining of artificial intelligence (AI) tools such as ChatGPT with traditional medical practices is becoming increasingly salient. Like mentoring a fellow or resident, the specificity, clarity, and nuances we convey to a generative AI system deeply influence the quality and relevance of its outputs.In this transformative era, the union of AI and oncology is not just a fleeting innovation but a paradigm shift, marking a new chapter in patient care. The intricate interplay of technology with the age-old wisdom of medicine has the potential to augment clinical decision making and broaden our horizons. The intricacies of oncology, with its ever-evolving landscape of treatments and methodologies, stand to benefit immensely from the precision and vast knowledge base of AI platforms such as ChatGPT.Yet, as with any tool, its utility hinges on its wielder's skill. Models such as ChaptGPT-3.5 cannot access real-time data (and have a cutoff date of September 2021), although newer models have evolving real-time functionality. By mastering the art of prompt engineering, oncologists can navigate the vast sea of information with precision, ensuring they extract the most pertinent, up-to-date, and clinically relevant insights. Such adept use of AI tools does not just offer more information, but it also promises clearer insights, richer understanding, and a more personalized approach to patient care, amplifying the efficacy of clinical decisions and thus optimizing patient outcomes.This new series highlights the mastery of “prompt engineering” in an oncology setting, aiming to enhance the harmony between clinicians and AI, ensuring AI is a valuable partner in delivering patient care.Ten Standardized Steps for Prompt Engineering in OncologyHere we list 10 key guidelines (in no particular order) to assist users get the best out of their AI queries: 1.Purpose DefinitionClearly articulate your objective and be specific. Instead of a generic “Tell me about lung cancer treatments,” opt for “Detail the latest advancements in targeted therapies for epidermal growth factor receptor-positive metastatic non-small cell lung cancer.”2.Use Contextual InformationEquip the AI with the setting. Instead of “What are common side effects of radiation?,” provide “Discuss the side effects of whole brain radiation in a metastatic breast cancer patient with cerebral metastases. My audience will be a patient with metastatic breast cancer and her family who are not AI experts.” Also make use of previous threads or even conversations earlier on in the same thread to provide further context for prompts.3.Structure the QuestionShape the output's format. To delve into surgical options: “Provide a step-by-step guide on performing a thoracoscopic lobectomy in early-stage lung cancer.” You can also specify the structure of the desired output. For instance, “List the top 10 most common radiation oncology CPT codes in JSON format.”4.Limit ScopeRein in your request to limit scope. For drugs, say: “Detail only the second-line immunotherapies used in advanced melanoma, excluding first-line treatments.”5.Control Output Length and FormatSet a number of words or characters for your desired output. Most modern models do not always provide answers within these limits, but they can still help with specifying your desired output. For instance, “Summarize this paper in 250 characters.” Provide examples in your prompt on the desired format of your output.6.Specify the Desired ToneDetermine the emotional undertone. For a patient memo: “Describe the potential impacts of palliative whole brain radiation in an empathetic, reassuring, friendly and easy to understand tone.”7.Role PlayAsk the generative pre-trained transformer (GPT) to play a role to better contextualize the intended response. Ask, “Act as a machine learning engineer and explain the top 10 prompt engineering tips to an oncologist.” Or “Act as a nutritionist and explain dietary and nutrition goals to a patient who will be undergoing chemoradiation therapy for a tonsil cancer.”8.Safety and Fact CheckingCross-reference AI guidance with trusted oncological resources, especially when discussing treatment dosages or new therapeutic modalities. If you are having difficulty understanding an answer, ask for specific examples.9.Explicit Over Implicit, Open Over ClosedAssume no inherent AI knowledge of your patient. For instance, specify, “Outline a treatment plan for HER2-positive breast cancer in a post-menopausal woman with a history of osteoporosis.” However, keep prompts as open ended as possible. Rather than “Is exercise important for reducing the risk of endometrial cancer?” try, “How does regular physical activity benefit patients with endometrial cancer?”10.Iterate and RefineDespite the most meticulous engineering of your prompt, you will likely still get an initial explanation that seems too general or too technical. Refine the question with follow-up prompts that help to continue the conversation or even challenge the GPT to list pros and cons of a situation before answering. For instance, if a GPT responds in general terms about resistance to poly ADP ribose polymerase (PARP) inhibitors for ovarian cancer, recalibrate with: “Break down the mechanism of action of PARP inhibitors in BRCA-mutant ovarian cancer for a patient newly diagnosed.” Engage in iterative discussions with the AI. If a treatment plan seems too generic, refine it: “Suggest a more personalized treatment regimen for a patient with triple-negative breast cancer and liver metastases.”Prompt AssistanceIn Figure 1, we provide an actual example of an AI prompt used in practice. This query was submitted to Chat GPT-4 on Wednesday, September 13, 2023. The response is provided without edits (Fig. 1).Fig. 1. Draft letter generated by ChatGPT in response to a real-world query posed by one of the authors (D.F.) in September 2023.PET/CT Scan Appeal for a Patient with Advanced NSCLC:“I need a comprehensive letter to appeal an insurance denial for a PET/CT scan. The patient, a 56-year-old male, has advanced non-small cell lung cancer (NSCLC). After a six-month regimen of combination chemotherapy, it's vital to assess disease status and progression. Due to a previous severe reaction to iodinated contrast, standard CT with contrast is contraindicated. The letter should be thorough, professional, and make a compelling argument for the clinical necessity of the PET/CT in guiding future treatment decisions.”ConclusionThis issue's example of a ChatGPT-generated insurance appeal letter is just one example of how generative AI tools can be leveraged to reduce administrative burden by speeding up the process of letter generation while still utilizing precise and detailed language to convey the intricacies of a common clinical scenario.As we embark on this journey together, we invite our readership to participate actively in shaping the future of AI in Precision Oncology. If you have explored innovative methods of prompt engineering or have suggestions to enhance AI–clinician interactions, we would be thrilled to feature them. We are excited to introduce a “Prompt Assistance” segment in our upcoming issues, and we encourage you to submit your ideas and experiences.Address your submissions to the editorial office of AI in Precision Oncology: [email protected]. Let us collaboratively push the boundaries of what is possible, ensuring AI serves as our ally in the noble pursuit of advancing oncological care.FiguresReferencesRelatedDetails Volume 0Issue 0 InformationCopyright 2023, Mary Ann Liebert, Inc., publishersTo cite this article:Douglas B. Flora and Nikhil G. Thaker.Designing Prompts for Generative Artificial Intelligence in Clinical Oncology Contexts.AI in Precision Oncology.ahead of printhttp://doi.org/10.1089/aipo.2023.0004Online Ahead of Print:October 17, 2023PDF download

Ähnliche Arbeiten