Obligatory—this is not medical advice
I know, perhaps I’m a little late into the fray here. I’ve had a few conversations with patients regarding the specter that is machine learning and artificial intelligence. ChatGPT is an interface that allows anyone to access a large language model.1 It is produced by OpenAI. This language model has been honed through the utilization of Reinforcement Learning from Human Feedback—alright, what does that mean exactly? The implications are that a user can use this as a tool. In essence, and this is my conjecture, you have multiple processes at work:
The ability to query a very large dataset incorporating history, science, mathematics, etc
A ‘reasoning’ capability that shapes that information and contextualizes it
The capacity to communicate this information back to the user in an interpretable form
This is a tremendous advancement. It basically gives people an automatic ‘expert’ in many domains. This is not to say that ChatGPT is always right, but its information is shockingly accurate most of the time. DALLE is, another model which uses similar methods to produce images.2 I couldn’t get it to draw a satisfactory picture of a soft tissue sarcoma. This could be user error.
A Conversation with ChatGPT
Let’s start broadly speaking, I asked it the following question:
This is honestly a great response and very close to what I have told patients in the past about their diagnosis.
There are some treatment plans that are less controversial than others, for instance low grade tumors.
Apologies here for lacking many details in the question—understand that I have gone back and forth with ChatGPT, and this is nearing the best case of its output. ChatGPT is right in broad terms—for easily resectable cases we remove the lesion with clean margins and discuss follow-up.
I proceed to name a more controversial case. I ask it about a 10 centimeter, high grade, pleomorphic sarcoma of the lower extremity—the model gets a little wacky if you name particular anatomic aspects, so these are eliminated here.
I would say that this is mostly correct. What is missed here is the subtle aspect of neoadjuvant vs adjuvant treatment, namely the possibility that radiation and chemotherapy could happen prior to a surgery. That stated, this is not the incorrect plan, particularly as I had mentioned that this is readily removed at the time of diagnosis. It even alludes to the fact that chemotherapy depends on multiple factors.
ChatGPT is also impressive at naming medications that *could* be used for treatment, although the context may be lacking. Specific combinations of treatment are not referenced (although there is an allusion to Adriamycin, ifosfamide, and mesna in the ifosfamide bullet).
I now ask it about a specific, ultra-rare, subtype of sarcoma and its FDA approvals. The data here were mixed, and I’ll mention why after we read their response.
Let’s break the response into individual components:
Crizotinib is FDA approved for patients with ALK-positive inflammatory myofibroblastic tumor
The mechanism of action as described here is correct.
The initial FDA approval date, in 2011, referenced is for ALK-positive non-small cell lung cancer
The phase II trial referenced here is for patients with ALK-positive non-small cell lung cancers3
The side effect profile here is accurate
The last portion noting that this should be done at an experienced center, and that clinical trials may be good options is also, as always, correct
What are my takeaways?
This is an incredibly powerful tool that will continue to improve. It is dramatically close to being able to answer these questions in a thought provoking and detailed way. I think that it’s only a matter of time prior to this serving as an important resource for our patients as they prepare for treatment—perhaps alleviating some of the educational aspects of healthcare. Likewise, it becomes easy to see how this variety of model could be integrated into other, hidden, aspects of the healthcare system. With a few tweaks, this could be employed to write notes, approve or disapprove of treatment plans, make nuanced surveillance recommendations, schedule and/or prioritize appointments/scheduling, etc.
https://openai.com/blog/chatgpt
https://labs.openai.com/
https://aacrjournals.org/clincancerres/article/20/8/2029/78791/U-S-Food-and-Drug-Administration-Approval
Fascinating!