Dear all, it is with great pleasure that I make public my latest exploration of openAI APIs. On this prototype, I have tested a medical chatbot.
Hope you enjoy the reading!
#bioinformatics #healthinformatics #medicine #chatbots #largelanguagemodels #openai #computervision #deeplearning #medicalimaging
https://www.medrxiv.org/content/10.1101/2023.12.31.23300681v1
You can leave a public review on
https://www.qeios.com/read/F9IVS2
Good job and thanks for sharing! My concern is related to the false positive and false negative outputs. After applying the app, does the medical doctor confirm all results by analysis/scan ect... How does it work?
Dear Loubna Youssar , thanks for the reply.👌
You are totally right regarding "false positive and false negative": this is a very important concern. For instance, on the paper, there is an example where the model marked viral pneumonia as urgent, and it is not urgent. What I believe one should pay attention is when an urgent is marked as non-urgent.
In my head, I would not eliminate the medical doctor, or a medical professional to double-check any diagnosis. I see this tool as a companion, not a medical replacement.
I believe the best is having a second level of checking. For instance, after the system created an appointment, a professional could check the case when possible.
My concern would be the serious cases, the non-serious sent as serious would most likely just waste resources.
Even though I am proposing this chatbot, I know it not trivial working this on a real-scenario. There are so many variable when it comes to healthcare.
My view is on case of a possible implementation, working closely with medical doctors and getting feedback. And adjusting the system to fit their needs.
Feel free to let me know more about your thoughts!💕👌
Yours Sincerely,
Jorge,
I believe that the serious cases will be treated regardless; however, those who test negative may still be sent home, even if they are unwell. I suppose the AI approaches alone are still far from being widely applied for several reasons. The primary concern is the issue of taking responsibility and data privacy.
For the next years, AI will be trained in parallel to the classical diagnosis till a full approval.
P.S: For those not familiar with AI in medicine. While AI has made significant advancements in recent years, its application in medicine dates back several decades such as X-ray, MRI, drugs and treatment therapies ect...
Dear Loubna Youssar
In fact, my background is on mathematical models, which faces similar issue. We have similar discussions on black box models, which tends to be more reliable.
"I suppose the AI approaches alone are still far from being widely applied for several reasons." indeed.
As I see it, most of the limitations are ethical, and acceptance. We already have models that arrive to human's intelligence. When we talk about AI mistakes, we forget that humans also make mistakes, and a lot. Medical doctors also make mistakes. On the main paper I have used, they compared the AI with medical experts, even though the AI did not surpassed all experts, it surpassed some of them, and kept a predictable behavior, less variance.
Indeed! It appears recently that AI has surpassed medical staff by achieving a 100% diagnosis accuracy for autistic children.
Dear Loubna Youssar , there are several cases, and they are not hard to replica as I have shown on my preprint.
It is similar to the self-driving car: it is no longer a technology limitation, rather a question of time.
Best regards,
Jorge
- Badges
- Science topic
- Similar topics
- Biological Science
- Immunology
- Antibodies