In new research, the tech giant said the Microsoft AI Diagnostics Orchestrator (MAI-Dx0) can correctly diagnose up to 85% of the New England Journal of Medicine’s case proceedings. That rate is more than four times higher than a group of experienced physicians. The best-performing MAI-DxO was paired with OpenAI’s o3 model.
Is the above true. We know a lot of hooey is in medical journals by Phara bought doctors and researchers.
I expect private equity to push “We don’t NEED no stinking human doctors!” meme
Of course there will some deaths; that is the cost of business and progress. And JCs get to remove high cost physicians from the cost equation.
Bottom Line! Profit!
Users shouldn’t say goodbye to their doctors just yet. “Your physician will be there to make the right judgment call at the right time, to plan treatment after diagnosis and to oversee and hold the AI accountable,” Suleyman said.
Soothing words as they ease humans out the door.
Are you vying for the Cynicism Award with @steve203?
The fact is that many (if not most) doctors are shoehorned into diagnosing and treating patients in 15 minutes. I know these people – they were classmates when I was pre-med. Most of them are smart but they are human. There is no way that they can work as doctors, stay up-to-date on medical literature and also have lives of their own.
I do a lot of research on my specific issues but I know that focus is only possible because I have all the time I need. I often say to doctors, “You have 1,000 patients but I only have one ME. And I’m a professional researcher.”
If a patient presents with a well-known problem the doctor can diagnose it and treat it readily enough. But there are many patients who present with unusual and/or combination problems. The symptoms can be common to several different disorders and it’s hard to sort them out.
One of my favorite books is " Every Patient Tells a Story: Medical Mysteries and the Art of Diagnosis," by Dr. Lisa Sanders. This was based on the continuing series of New York Times columns by Dr. Sanders. There was also a TV series with real patients.
The patients have been previously seen by multiple doctors who were unable to diagnose the problems, which range from rare genetic diseases to tick-borne fevers to overdoses of Pepto-Bismol.
I have already used AI (Google Gemini, ChatGPT and Perplexity) to study DH’s case which included pulmonary nodules plus emphysema. (Fortunately, biopsy showed the nodules to be benign scarring, not lung cancer.) The AI answers were detailed and linked to research references.
I am really excited about a diagnostic AI which has been designed from the start for medical consultation. There is no way that this will cause doctors to be fired. It will help them diagnose and design treatments based on the latest research.
I am not cynical about this any more than I was cynical about the introduction of spreadsheets displacing accountants.
The goal of the AI is clearly diagnostics – that’s what it is designed to do. What executives do with the capability has nothing to do with the AI itself.
Wendy
Both can be true. I seldom make predictions, but I will this time: AI is coming for lots of jobs, and fairly soon too. I used to think my field (environmental consulting) could never be outsourced or automated. I no longer think that. It is easy to see how AI could replace the large majority of my work product. It hasn’t yet, but it will. I’m mostly retired but take on selected projects from time to time. I’ve been using AI to increase my output for certain tasks, and it has helped a lot. I’ve mentioned this in private industry discussion groups and people seem surprised by this. There are certain task AI does not yet do, or doesn’t do well, but I’m 100% certain this will change.
That means fewer consultant hours will be required per project. That means fewer consultants will be required, period. The Zuck recently said that AI is a good as a mid-level programmer. That means no need for human mid-level programmers. If not now, then soon.
In one sense, maybe this is like Excel making workers more productive. Instead of hiring a mid-level programmer, I can now have AI do it for me, increasing my work output. I dunno, though. I got a bad feeling about this one.
These days PCPs wish they only had a 1000 patient panel. Last I heard it was more like 2000. Some docs move to Direct Primary Care or concierge practices so they can have smaller panels, 300-500, and devote more time to each person. I think AI has alot to offer diagnostically. The one thing it can’t do is “care”. I still think educated human encouagement and guidance counts for alot.
Both “paranoia” and “cynicism” imply a degree of mental disorder leading to unrealistic fears. As with the paranoia case, is the person being “cynical”, or “realistic”?
You can become your own primary MD but it does take time and effort. While one is in the rat race one has little time for oneself. Suddenly you find yourself in the ER.
@captainccs I was pre-med in college with a double major in chemistry and biology. I scored in the 99th percentile on the Medical College Admission Test (MCAT). I didn’t go to med school because I don’t have the physical stamina to do 36 hour shifts which were required from interns. A 9-to-5 job as an industrial chemist was within my physical capabilities.
But diverting from med school didn’t stop my interest in medicine. In fact, it enabled me to explore my interests in preventive medicine, especially diet and exercise, which med schools don’t teach.
I’m able to understand research papers so I’m qualified. This doesn’t work on a Macro scale because most people don’t have the education. The benefit of AI, such as ChatGPT, is that the AI will translate research results into easier-to-understand plain language. ChatGPT even asks if the user wants a letter to bring to the doctor.
Wendy
Yes, very much so. Aside from having the right education and the requisite intelligence (enough to know about and use ChatGPT and other AIs that use extremely large language models) is simply knowing how to ask the right questions. One example is at The Art Of Prompts – Asking The Right Questions To ChatGPT | Devon Geography
Of particular importance is assigning the role of expert/authority in the category or field you are querying. Until reading through several of these “prompt” charts, I had not realized what a difference it can make in the answer you receive back.
Read a different article, buried in the story, only trained on about 300 cases. Didn’t see anywhere what kind of cases. Were the concentrating on common scenarios (horses) or presentations of rare diseases (zebras)*. I still revert back to what my BIL told me about AI tests they’ve run on super computers he help build. He gave the example of pictures of blueberry muffins and chihuahuas and can AI tell the difference. It eventually learned and could become fairly accurate on the subset of 1000 pictures, but give it a new subset and all bets were off. I can see where AI could be another tool but not replacing humans. Too many nuances that can’t be programmed.
An old medical saying of thinking about diseases in patients, when you hear hoofbeats, think horses not zebras. The comedian in me always asked what do doctors on the Serengeti think?
Have you ever tried to figure out the sample size a human experiences by the time he/she becomes a teenager? How many times 1000?
That was the case with algorithmic AI which failed miserably. Neural network AI uses an entirely different paradigm that relies on statistical probability, the larger the sample size the more accurate the output.
The Captain
12 years
365 days
16 waking hours
60 minutes
4,204,800 imprints?
@MataroPete, I think you’ve inadvertently given me the answer to my question of “How did he do it?”…my ex son in law, that is, in composing his Grand Opus designed to bamboozle the domestic abuse evaluator. There were a whole series of questions in one of the questionnaires that it’s totally possible to answer "in the style of " a long term victim of… and you fill in whatever your imagination can conjure up.
