Personal Science Week - 230511 Doctor GPT4
Use GPT-4 for medical-grade health and diet advice, plus book reviews and more links
This is a weekly update for anyone who wants to use science for personal, rather than professional reasons.
This week we continue our discussions of AI in healthcare with some new results from GPT-4.
Personal Science is delivered free each Thursday to anyone who wants to be a Personal Scientist.
Back in Personal Science Week - 230316 AI Chat we showed an example of how anyone can use ChatGPT for a basic blood test analysis. Since then, much has happened and we were able to test against the new, improved GPT-4 system with impressive results. Here’s how to do it yourself:
In all of the following examples, my questions (“prompts”) are in bold and the answers are in italics.
The main rule is to start with as much detail as you can. I made up the following example to give you the basic idea. (Personal Science paid subscribers can see the entire 15-page example of my actual personal report)
You and I are medical doctors discussing one of my patients. He's a 45-year-old male with no family history of heart disease, diabetes, or dementia. His paternal uncles have had strokes before age 70, and his grandmother died of breast cancer at age 62. He rides his bike to work about 5 miles each way daily on mostly flat terrain. His blood pressure is 120 / 70. Weight 190 pounds, 6 feet tall. Non-smoker. Drinks about 2 glasses of wine each weekend. Sleeps about 7 hours per night, more on weekends. Omnivore, eats red meat most meals, with two servings of raisin bran and milk and a glass of orange juice each breakfast. He complains of annoying constipation, generally only two or three bowel movements per week for the past several years. In my next prompt I will copy/paste the results of his most recent blood test.
Thank you for providing the patient's background information. This is helpful in better understanding his overall health and identifying any potential risk factors. It's good to see that he has some positive health habits such as biking to work and being a non-smoker. However, there are some aspects of his lifestyle that could potentially be improved, such as his diet and alcohol consumption.
I'll be able to give you a more informed opinion once I review the results of his blood test. Please go ahead and provide the blood test results in your next prompt.
Add Your Own Data
Next, open the web site where you keep your blood test results. If you received it as a PDF, just select-all and copy/paste it into the prompt. The system will respond with a point-by-point analysis like this and a conclusion like this:
Based on these results, your total cholesterol, LDL cholesterol, and apolipoprotein B are elevated, which might be a cause for concern.
I have the patient’s exome SNPs as well. Could you propose a list of SNPs that might be relevant to his situation. Please return in table form, showing just the recommended genomic results to study further. Include the RS numbers (with 23andme URL if possible) and a short summary of current research on those SNPs
GPT-4 will respond with a table, including URLs you can click to go straight to the 23andme web site. If you already have an account with results from 23andme, you can log in and see the specific SNPs one at a time. Then simply copy/paste them into the prompt
Here are my SNPS: rs1800562 | A/G rs1799945 | C/C rs4149056 | T/T
Consumer genomics test usually return “rs” identifiers you can think of as the location for various gene variants. At this point, you’ll probably want to proceed interactively. For each rs value, GPT-4 will have some information about the implications for this patient. Depending on what it returns, you may want further details.
Now, taking these genotype implications in mind, please redo the summary of the patient's visit, and be specific about personalized recommendations.
When I did this on my data, GPT-4 returned a table like this:
Make Diet Recommendations
Now generate a dietary recommendation. Suggest overall macronutrient percentages, including fat, protein, carbohydrate, plus fiber and sugar. Make a chart with recommended daily percentages of each category, a brief reason for why your recommendation is different or specific to this individual compared to average, a few specific foods you recommend he avoid eating. Only name foods that might ordinarily be considered healthy, but should be avoided for this specific person. If you can’t think of any, say “N/A”.
Based on the patient's genotype and blood test results, here are some specific, otherwise nutritious foods that this patient may consider avoiding or limiting:
Get Precise Menu Advice
Now suggest 3 days worth of meals, assuming the patient is an omnivore who lives in the Pacific Northwest and enjoys a variety of ethnic cuisines, especially Chinese (Yunnan), Japanese, Mexican (Oaxaca) and has access to specialized ingredients as necessary. Be sure to indicate volumes for each meal. The patient drinks coffee regularly with cream, but otherwise doesn't have snacks or alcoholic beverages.
GPT-4 responds with a lengthy 3-day menu. Here’s a typical dinner:
Having experimented with this for several weeks on many different kinds of data, I’ve concluded that there is no “perfect prompt”. The best way to get medical advice from GPT-4 is interactively: ask questions, consider the answers critically, and then ask additional questions. This is not unlike the way you should talk with a human physician.
AI Revolution in Medicine: GPT-4 and Beyond by Microsoft’s Peter Lee and team goes into much more detail about the impact GPT-4 will have on doctors.
The authors of this book should really be editors, since much of the text is taken verbatim from interactions with GPT-4. That's not necessarily a criticism -- a subject like this begs for specific examples. A book that describes how to use Excel, say, or some programming language, will consist mainly of text directly output from the computer.
But they go a little too far when, for example, in Chapter 9 ("Safety First") they simply create two personas and ask GPT-4 to generate a discussion. The results are interesting and provocative, but I left feeling, well, like why did I need the authors for this? Wouldn't I be better off just asking GPT-4 these questions myself?
That question is a good microcosm of the whole issue of using GPT-4 in medicine. Did I really need these authors to go to the trouble of making a book? That's eerily similar to the question of do I really need a doctor to tell me that these symptoms are congenital adrenal hyperplasia (one of the GPT-4 generated cases they present). How much will I really need doctors now that I have GPT4? Editors after all are useful, to help direct our thoughts in a particular, useful direction. Maybe in the future we'll think the same of doctors.
If you want a good summary of how GPT-4 will change medicine, a more concise summary is in the one hour YouTube talk by author Peter Lee: Emergence of General AI for Medicine. Or see the Special Report they published in the March 2023 issue of New England Journal of Medicine. (available as a PDF for free if you register).
More AI-related links
This Feb 2023 study, “Performance of ChatGPT on USMLE: Potential for AI-Assisted Medical Education Using Large Language Models” concludes
ChatGPT performed at or near the passing threshold for all three exams without any specialized training or reinforcement
Paradoxically, ChatGPT outperformed PubMedGPT (accuracy 50.8%, unpublished data), a counterpart LLM with similar neural structure, but trained exclusively on biomedical domain literature.
The authors speculate that this is because general purpose models tend to come from large pools of what amounts to consensus understanding, whereas specialized data tends to include too many exceptions, often contradictory.
A one-hour NIH lecture shows what happens when we measure one person longitudinally instead of after symptoms appear. This proteomics test shows something weird happened more than 1 year before the bladder cancer diagnosis.
About Personal Science
AI and other technologies are moving so quickly that it’s hard to keep up. Personal Scientists like to stay on top of things, of course, but sometimes it’s better to take a longer, more skeptical view. What is the scientific method, really, and how does it apply to personal questions?
If you have ideas or other topics you think will benefit other Personal Scientists, please let us know.
If you have a lot of results, you can also ask GPT to generate a Python script that will do this automatically. It took me only a few minutes extra, but the details are beyond the scope of this short newsletter.