I envy medical students and residents who will be able to focus on patients without constantly toggling between faces and a computer screen. It’s why I dove headfirst into generative AI and volunteered to write this white paper, hoping to help other clinicians, researchers, rare disease advocates, and healthcare professionals. This is a perfect primer for those who don’t know what ChatGPT is as well as those who have only dabbled but want to learn more.
What is Generative AI and Why it Has Healthcare’s Attention
So what exactly is generative AI? Why are we writing about it now? Simply put, generative AI refers to artificial intelligence systems focused on the actual creation of new content, rather than just analyzing existing data. Thanks to progress in deep learning and large language models over the past decade, AI systems can now generate amazingly coherent text, images, audio, and more from a sentence. In fact, generative AI wrote this paragraph! This came to a head with the 2022 launch of OpenAI’s ChatGPT and has since exploded into every facet of our lives. ChatGPT has over 100 million users, billions in revenue, and even has its own soap opera.
What Makes Generative AI Different from Past Forms?
AI platforms use a technology similar to autocomplete that keeps on going, adapting what it creates to the context. It can help ideate, expand on concepts, rewrite poor phrasing, and more. Generative AI can function like an assistant, understanding instructions delivered in plain language, or natural language in the industry parlance.
AI Models Comparison: How is it different
Generative Nature
- Past AI Models: Rule-based, lacked content generation.
Example: Followed specific rules for diagnosing diseases. - Generative AI: Creates new content based on learned patterns.
Example: Analyzes diverse patient data to generate new insights and recommendations.
Pre-training and Transfer Learning
- Past AI Models: Task-specific, required extensive rules.
Example: Required specific programming for each task. - Generative AI: Pre-trained on large datasets, adaptable to various tasks.
Example: Can carry out tasks with a simple natural language prompt like “what disease is most common in my patients?”
Scale and Data Dependency
- Past AI Models: Limited by computing power and data.
Example: Struggled analyzing GBs of images for dermatologists. - Generative AI: Large-scale models like GPT-3, effective with vast datasets.
Example: Much more accurate diagnosis of radiology conditions.
End-to-End Learning
- Past AI Models: Modular approach with separate components.
Example: Separate modules for patient history and symptoms etc. - Generative AI: End-to-end learning, single model for tasks.
Example: One model for everything.
Natural Language Understanding
- Past AI Models: Struggled with language nuances and context.
Example: Misunderstood subtle nuances in patient reports. - Generative AI: Improved understanding, generates coherent responses.
Example: Interprets complex medical narratives without new training.
Transferability of Knowledge
- Past AI Models: Task-specific knowledge, lacked transferability.
Example: Train a model for each specialty. - Generative AI: Transfers knowledge across tasks, adaptable.
Example: One model trained on dermatology can be quickly adapted for cardiology.
The Unique Powers of Generative AI for Healthcare Users
Generative AI is moving at light speed, it can be overwhelming to keep up. We will start by focusing on how my colleagues and I personally use generative AI in a healthcare setting, then transition to how clinicians, researchers, and advocacy groups can use it at the clinical and corporate levels.
Note their prediction: ChatGPT alone is at over 100 million users in under 1 year.
How a Medical Professional Uses Generative AI Daily
As a practicing doctor, published researcher, and Substack writer, I have woven these emerging tools into my workflows, and I highly recommend you do the same.
Superpowers and Tools
Summarization
- What I use: ChatGPT3.5, Claude, Bing (Uses ChatGPT3.5 and 4)
Knowledge management
- What I use: Paid subscriptions to ChatGPT and Claude allow you to upload PDFs that you can summarize.
Generating text
- What I use: Claude for long writing. It handles longer texts more effectively because it has a longer context length.
Generating images and video
- What I use: Bing, Midjourney (paid) for really beautiful images. Has a learning curve, but their community is quite helpful.
Generating code
How a Physician Uses Generative AI: Scribes, Coding, and Billing
Physicians love talking to patients, getting to know them as people, and solving their problems. Having a computer between a physician and patient is a miserable experience. The documentation burden during the visit is a problem, while every clinician knows the amount of typing that must be done after the patient leaves. It's a major determinant of clinicians leaving medicine or as Dr. Sinsky put it, “Physicians don't leave their careers. They are leaving their inbox.”
Many clinicians have started using digital scribes, which “listen in'' on conversations and put that information into note form. These tools have been revelatory for many clinicians, allowing them to recapture their love of medicine and spend more time fully engaged with their patients. Though I haven’t gotten to use these tools yet in my practice, I’ve tried multiple models and expect these to be standard within the next few years.
My practice also uses AI-enabled software for mundane administrative tasks like coding visits and tracking billing details. Most physicians would prefer never to think about how to bill or code patient visits ever again, and I am hopeful that time is coming soon in my practice as well.
How a Researcher Uses Generative AI
For manuscripts I author, generative AI makes the process of reviewing what studies have been done much faster. Previously, I would type in a variety of possibly relevant search terms, read each abstract, decide which articles were worth reading, and summarize that information. Generative AI expedites every step by allowing me to input questions in natural language, giving me an overview of the most important and helpful articles, and writing a brief literature review. This accelerates the writing process so I can focus on interpreting the studies and synthesizing the information into new insights.
Even better, I often ask AI tools to give me feedback on my manuscripts before I submit them. I ask them to pretend they’re an editor of a medical journal, and they give surprisingly good critiques and suggestions about what to add and where the manuscript could be more clear.
How a Healthcare Influencer Uses Generative AI
Writing is where I use generative AI the most in my daily life. I write a weekly Substack newsletter for physicians about healthcare AI and use AI tools for almost every part of the process. I usually write an outline of my planned newsletter and ask generative AI if there are any major topics I’ve missed. Then I start to write, and if I run into writer’s block, I ask the AI platform for a sample paragraph to get me started. Often I use only a sentence from what the platform wrote, but it’s enough to get me going. This process lets me focus on providing unique insights instead of the operation of writing.
Once I complete the piece, I provide AI text from each section and ask it to create an image based on the text. The image creation feature is pretty incredible; sometimes I’ll choose a favorite artist and ask them to make all the images in the style of Degas or Frida Kahlo, for example.
How to Use Generative AI to Analyze Data
I have spoken with audiences as varied as rare and orphan disease advocacy groups, regional health insurers, and public-private partnerships, and they all want to know how to use generative AI to make some sense of their data since they don’t have a large data science team. I often use generative AI for help with Excel and SQL queries when analyzing data. One major leap I’ve made is utilizing generative AI to write code to pull information from a website and analyze it. For example, I will ask it to give me code that pulls all articles from a website and identify the top five most mentioned clinical terms.
How I use Generative AI in my Personal Life
Family
- Create weekly meal plans and grocery lists
- Find the most fun activities for kids in any random city in the world
- Figure out how to do household projects
Clinical
- Maximize efficiency for scheduling patients
- Develop a monthly call schedule given vacation constraints
- Review AI-assisted clinical decision support
Research
- Review references and summarize existing literature
- Write a first draft from outline form
- Edit and reformat into different submission types
- Critique drafts
Business
- Use AI-powered calendars to group my meetings efficiently
- Use AI voice summarization tools like Otter AI to summarize my meetings and find facts after the meeting is done
- Create images for flyers and promotional materials
Leveraging Generative AI to Transform Medicine
I am thrilled by the prospects of generative AI across healthcare as a whole. Drug discovery, medical imaging, personalized medicine, and research analytics are already being transformed to improve diagnosis, treatment, and patient care.
Drug Discovery and Development
The average time between a drug being discovered and being widely available is 17 years, a full generation. Generative AI promises to accelerate the traditionally lengthy, expensive, and inefficient process of discovering and testing new medications. Its unparalleled ability to analyze associations within massive biological datasets enables the rapid identification of novel drug candidates.
Powerful generative models can predict how molecules might interact within the body and suggest chemical tweaks to optimize how well the drug will work in the body. AI tools can also make it faster and easier to recruit patients for clinical trials, further decreasing the time to FDA approval.
Google’s AlphaFold heralded a new era in the prediction of protein structure and solved problems in minutes or hours that used to take months or even years. This breakthrough significantly advances our understanding of biological functions and facilitates drug discovery.
Medical Imaging Analysis
Applying generative AI to analyze medical images has shown particular promise for earlier and more reliable disease diagnosis. Image-heavy fields like radiology, pathology, and dermatology have incorporated AI more quickly than other fields, and most of the AI algorithms approved by the FDA are related to radiology. The reason is that there are substantial existing datasets to train against. Smart algorithms act as a “second reader” for radiologists, providing another layer of screening to catch minor abnormalities. Sometimes AI tools can even pick up patterns that are too subtle to be seen by the human eye, like patterns in EKGs that predict kidney failure or differences in the retinas of women and men. This ability has especially important applications in rare diseases, in which previously unrecognized patterns may hold clues for future advances.
Personalized Medicine
The great joy of medicine is that each patient is unique, unlike an airplane or computer. That variability makes diagnosing and treating individual people more challenging since each patient may have slightly different symptoms and responses to treatments.
Research teams have already demonstrated success applying generative AI to predict cancer patient responses and tailor individualized therapeutic regimens integrating genetic data, past outcomes, and cutting-edge research. Twenty years ago it would have seemed miraculous to test a patient’s cancer cells to determine whether she will respond to a medication, but now it’s routine. Such hyper-personalization will only expand as predictive algorithms grow more powerful.
Especially in rare diseases, generative AI can enable truly personalized medicine by clarifying these complex connections within patient data. Sophisticated models could decrease time to diagnosis by helping physicians consider rare diseases earlier in the course, as well as personalize treatment plans.
Healthcare Administration and Research
In addition to using generative AI to help patients directly, this technology can decrease the time physicians, nurses, and others in the healthcare team spend on tasks like scheduling, billing, and paperwork. Innovations affecting patient care are higher risk, so it’s no surprise that AI tools decreasing the administrative burden have been some of the first adopted. These tools allow clinicians and administrators to spend more time with patients and, ironically, away from their computers.
The amount of data that clinicians and hospitals collect now is staggering compared to only a decade ago. However, the data is complex and difficult to analyze and has sat unused on hospital servers. AI tools promise to sift through these massive silos of patient data to advance understanding and help researchers develop insights. This ability to analyze and organize large datasets will be especially beneficial for rare diseases, as previous data analyses may not have been adequate to identify incidence rates, contributing factors, or potential treatments.
Generative AI as an Equalizer
As a physician who has worked with both large hospital systems and small private practices, I’ve seen first-hand how resources and capabilities can vary wildly. By leveraging text and data-focused AI systems, small groups can effectively “punch above their weight” to generate volume and sophistication previously only achieved through significant human effort. This applies to rare disease advocacy groups trying to spread awareness and educate on overlooked conditions as much as small private practices striving to provide high-quality care without the budgets of neighboring health systems. I’ve also spoken with people who have rare diseases who’ve mentioned the challenges of trying to wear many hats while dealing with few resources. Generative AI can take on tedious tasks and allow groups with limited numbers to increase their impact.
The Underdogs and AI Amplifying Effect
Rare Disease and Orphan Disease Advocacy Groups
- Finding and communicating with members
- Creating content
- AI models testing new cures
Small Physician Groups
- Scribing
- AI patient engagement
- New patient marketing
Regional Health Insurers
- Access to powerful AI models for risk mitigation and stratification without internal data science teams
Researchers
- Summarizing massive datasets and literature quickly
Let's walk through some tactical examples of how generative AI can serve as an equalizing force multiplier. Content creation is one major use case. It’s now almost instant to draft blog posts, social media captions, newsletters, and other promotional material to raise awareness. An AI assistant takes high-level messaging guidance and turns it into polished copy customized to a specific audience and voice. This allows understaffed rare disease and orphan disease advocacy groups to spread understanding of overlooked conditions at scale, for instance.
Consuming and summarizing relevant medical literature and emerging research is another invaluable application. An AI platform can rapidly digest large quantities of studies then create summaries and insights based on that information. Perfect for physicians and nurses at small practices trying to keep up with the latest in clinical research.
Generative AI also accelerates essential administrative tasks like identifying prospective partners and contributors. With thoughtful prompts, AI can efficiently scrape databases and web content to generate targeted lists of medical directors aligned to collaborate on clinical trials or potential donors to contact about your cause.
Utilizing open-source generative AI can help regional health insurers automate patient engagement without having a large dataset to train against. This could be the ultimate democratizing force in healthcare - giving more groups, regardless of size and means, a fighting chance to drive progress with technology's help.
How to Use Generative AI
I’ve told you a lot, but how do you use it?
Getting Hands-On with Leading Generative AI Tools
There’s an old adage encouraging medical trainees to "see one, do one, teach one." While we’ve gone away from that in patient care for safety reasons, the saying is very applicable to generative AI. Practical first-hand use of these tools will give you a much better understanding of their capabilities and limitations. Each type of task, like questions, image generation, and coding, has a different tool and approach.
Questions
I ask ChatGPT or Claude questions throughout the day about non-medical, daily life questions like trivia that my family is arguing about as well as practical questions like “create a four-day itinerary for a family of six for Florence, Italy that includes the major highlights plus some less common sights.” I’ll also ask for recipe suggestions that use the random four ingredients in my fridge. This kind of regular use has given me a much better sense of what some have called the jagged frontier of AI, in which it performs well at some tasks like those above, and terribly at others like anything involving numbers. It also gives step-by-step directions well, and in a much easier to follow format than a web search would do. If you need to complete a household project, it can find you a good video to watch and outline the steps.
If you are just starting out learning how to speak “AI,” OpenAI has a helpful best practices guide to check out.
Images
For visual media generation, I use Bing regularly. Bing has access to ChatGPT-4 (but only in creative and precise modes) as well as DALL-E, which is an image generation platform. Typing in a prompt produces shockingly good images. If you’re willing to put in some time learning how to use it, Midjourney excels at stylized interpretations and produces some really beautiful pictures. Images from these tools can help explain concepts through approachable graphics and metaphors in articles or PowerPoint presentations.
I asked DALL-E 2 to create an image of someone learning AI for the first time in the style of The Lion King movie from 1992 (note: it’s from 1994). Notice the colors in the bottom right corner denoting that it is an AI image.
Code
For automating workflows, adapting solutions to local needs, or building customized interfaces, tools like Co-Pilot on GitHub and Codeium translate natural language instructions into functional code across multiple programming languages. Friends who code regularly tell me that these tools are especially good at debugging code and identifying why the program is not working as expected.
Research
A recent survey of 1,600 researchers shows considerable optimism about incorporating AI into research, with more than 80% of early adopters saying they consider AI to be helpful. Studies in other fields suggest that humans paired with GPT tools come up with more innovative ideas for writing topics. I frequently use AI for brainstorming, organizing ideas into themes, and quickly summarizing existing research. New tools like Elicit and Scite find important papers on specific topics and even summarize them. Generative AI is also great at writing critiques. For example, you can ask it to pretend to be a donor and request 5 ways to improve the grant application.
Once you’ve figured out how to use AI, you will want to apply it to your business. You may not have the expertise in-house. There are open-source marketplaces where others share their healthcare AI models, and you can utilize and alter them to fit your needs.
Marketplace for Healthcare AI Models
The advent of vast model repositories like Health Universe and HuggingFace uniquely democratizes access to leading-edge AI capabilities. These platforms essentially function as generative model app stores, allowing users to quickly browse pre-trained models spanning from the Batch Charlson Comorbidity Index to medical image segmentation.
Patients, advocacy groups, and healthcare teams can repurpose or fine-tune models for entity recognition, clinical note analysis, cohort stratification, and more in just days. Most importantly, little to no coding experience is required to use the majority of these tools, making data analysis and research tools far more accessible. For example, Health Universe’s focus on rare diseases makes it especially exciting for patients and advocacy groups, as it allows them to benefit from tools developed for other disorders with a large comment section to share feedback with those who have already been through the process.
HuggingFace houses over 30,000 models with intricate tagging and performance benchmarking data to ease discovering optimal choices. And seamless integration options with popular data science workflows via Docker containers, Python packages, and cloud compute resources mean models can integrate into production within days compared to the months of effort conventionally required.
Between readily customizing pre-trained models and contributing their own iterations back to the community, these marketplaces embody the collaborative potential of AI to push scientific frontiers ever further. These tools democratize pioneering innovation.
What Generative AI Does Poorly for Healthcare as of Now
Generative AI is not perfect, so we want to illustrate a few limitations we’ve seen.
Hallucinations
Hallucinations refer to instances where AI systems generate misleading or incorrect information, images, or perceptions that are not grounded in reality. The main generative AI tools like ChatGPT, Bard, and Claude struggle with citations, whether from scientific literature or news articles. We have spent an embarrassing amount of time trying to find events or papers that don’t exist. Our recommendation is not to use generative AI to seek factual truth.
Writing Finished Products
Although generative AI can seem like a miraculous writer, it’s kind of the Muzak of writing. Inoffensive and easy to read but rarely engaging enough to keep your attention. In other words, it can write a first draft but not the last draft. We initially wrote this white paper using generative AI, and our eyes glazed over reading it. Our recommendation is to use it for outlines and inspiration, not the finished product.
Math
Due to the technical details of how generative AI works, it can barely do math at the level of a toddler. A calculator is miles ahead of it. Here is a nerdy explanation why, but this limitation will not last forever.
Evaluating and Choosing AI Tools
Generative AI technology has improved so quickly that it can feel overwhelming to find the right tool for your needs. Many of the considerations related to choosing generative AI tools overlap with your existing evaluation framework. Some new considerations around technical requirements, data compatibility, and information privacy and security are paramount with any healthcare AI tool.
Technical safeguards around access, auditing, and correction processes are foundational. But applying healthcare's guiding principle of "first, do no harm" means explicitly prioritizing ethical considerations like mitigating risk and bias and defining accountabilities across generative AI deployment, monitoring, and adaptation prove equally key to sustainable adoption.
Fortunately, the same patient privacy laws that have been in effect since HIPAA was passed in 1996 apply to generative AI just as it does for any other technology. And despite headlines that describe how generative AI will replace physicians, that simply isn’t happening. Healthcare is complex, and generative AI is a tool that should facilitate, not eliminate, the physician-patient relationship. Plus, human guidance plays a crucial role in spotting areas where AI may hallucinate or demonstrate bias rather than produce completely factual information. Our recommendation is to focus on generative AI tools that amplify organizational knowledge rather than replace it.
Deploying AI Tools in a Clinical Setting
A recent survey by Wolters Kluwer found that:
“When asked if they would be concerned knowing that their healthcare provider was using GenAI, four out of five (80%) respondents said they would be concerned.”
Be transparent and specific about where generative AI is used, why it is being used, and the safeguards in place. Be sure to communicate that the final work was approved by a researcher or clinician. Always check sources whenever receiving statements of fact from generative AI. Finally, make sure to communicate the positives that are coming out of it. For example, thanks to generative AI, we are able to spend 5 more minutes with each patient. This will make patients more tolerant when, like any new technology, there are issues.
