Artificial intelligence in medicine and science

[rvallee]

Conversational diagnostic artificial intelligence in ambulatory primary care: a prospective feasibility study

https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(26)01535-7/abstract

Background​

Artificial intelligence (AI)-based systems show promise for assisting primary care providers (PCPs) with patient care. We aimed to evaluate the safety and quality of clinical conversations of a patient-facing conversational AI system, which engaged in real-world urgent primary care appointments.

Methods​

In this prospective, single-centre, single-arm feasibility study, English-speaking patients aged at least 18 years interacted with the Articulate Medical Intelligence Explorer (AMIE) up to 5 days before a single-complaint urgent primary care appointment. Physician safety supervisors monitored all interactions and were trained to intervene on the basis of predefined safety criteria. AMIE transcripts and summaries were shared with PCPs before the visit. Primary outcomes were the number of supervised conversation safety stops, AMIE's conversation quality assessed by clinical evaluators, and patient and PCP experiences per surveys. This study is registered with ClinicalTrials.gov (NCT06911398).

Findings​

From April to November, 2025, 114 patients were enrolled with 98 completing both the AMIE interaction and the PCP appointment. Zero conversation safety stops were required on the basis of prespecified criteria. Safety supervisors noted one hallucination and added clinical information in five interactions. AMIE's conversations were rated favourably in 87–100% of cases (17 criteria) by clinical evaluators, and 48–96% (16 criteria) by patients. Patient attitudes towards AI improved after interacting with AMIE and remained elevated after the patient's visit with their physician. PCPs completed post-surveys in 60 of 98 cases, including 44 cases in which they reviewed the AMIE transcript before the visit. PCPs found AMIE helpful for visit preparation in 33 of 44 cases and reported that it might have changed their behaviour in 25 of 44 cases.

Interpretation​

Although further research is needed, this study shows the initial feasibility of conversational AI in a real-world setting—assessed via conversation safety and quality, as well as user acceptance—and represents a crucial step towards clinical translation.

 

[Creekside]

Having what is in effect your own personal physician available 24/7 might reduce delayed diagnoses (and therefore treatment or prevention advice). How many people notice possible medical problems, but put off seeing a doctor because of cost and inconvenience or embarrassment or fear of being thought of as a hypochondriac? I can think of several times when I should have gone to my GP much earlier, but didn't.

 

[rvallee]

Having what is in effect your own personal physician available 24/7

Think bigger: having an entire world-class hospital of physicians, all the specialties, all the sub-specialties, all of which are organized in an orchestra of verification and reasoning with the goal of solving the same problems. The best physician in the world, if that's a thing that can be rated, is nothing compared to that, even available 24/7. Medical AIs will have millennia of clinical and research experience on second one.

 

[nafione]

Think bigger: having an entire world-class hospital of physicians, all the specialties, all the sub-specialties, all of which are organized in an orchestra of verification and reasoning with the goal of solving the same problems. The best physician in the world, if that's a thing that can be rated, is nothing compared to that, even available 24/7. Medical AIs will have millennia of clinical and research experience on second one.

Do you work for OpenAI et al? You don't seem to apply your usual skepticism for all things me/cfs to machine learning tech.

 

[nafione]

Glass AI--AI that generates differential diagnosis and clinical plans

Recently I learned about this AI tool, which is currently available for anyone to use without cost or registration.

Edit: Darn I forgot the link: https://glass.health/ai

I'm playing around with it to see how well it understand ME/CFS. It seems like it's familiar with it. Here's the DDx it generates for one case I made up:

I don't think this is a meaningful test, if you feed the AI the clinical language directly related to me/cfs. My language has evolved considerably from pre to post diagnosis, and I still struggle to explain it to others.

 

[Creekside]

You don't seem to apply your usual skepticism for all things me/cfs to machine learning tech.

Well, AI isn't limited by self-serving humans, so there's more hope for good outcomes.

 

[Creekside]

Think bigger: having an entire world-class hospital of physicians, all the specialties, all the sub-specialties, all of which are organized in an orchestra of verification and reasoning with the goal of solving the same problems.

Yes, but you wouldn't interface directly with the sub-specialist. The AI GP would bring those others in when necessary.

We are so far avoiding paying full cost of AI service. It sounds like AI companies still in the "milk investors using wild promises" phase. Learning to be the friendly GP for an individual requires a different set of resources than diagnosing complex diseases, and with different size of requests, so referrals seem logical.

Hopefully the "gouge the customers to the max" phase will be short, and replaced by a completely different economic system.

 

[rvallee]

Do you work for OpenAI et al? You don't seem to apply your usual skepticism for all things me/cfs to machine learning tech.

No, but this is what I would be working on if I wasn't ill. It's just hard to understand how information sciences scale, how you can get several billion-fold improvements in performance. There's nothing else like it.

 

[rvallee]

We are so far avoiding paying full cost of AI service. It sounds like AI companies still in the "milk investors using wild promises" phase.

It's more that the initial costs are very high, but keep going down. Something like 50% per quarter over the last 4 years, or something like it. What you can do with $100 of a top-end model, at full cost, will be a lot cheaper and better in 6 months, then the same after that. And we don't even have ASICs yet, where parts of the models are baked in the chips.

The hardware industry saw similar scaling, although working purely with software adds a whole other layer of optimization. A current smartphone has more computing power than all the world's computers at some point in the past, probably around the late 70s.

Exponentials are wild, it's very hard to wrap your head around them, but so far nothing I've seen out of AI has surprised me. This stuff just works.

 

[hotblack]

It's just hard to understand how information sciences scale

So this is the thing that I think is far more likely than some superhuman intelligence. It’s human level intelligence or rather knowledge work as a commodity. The industrial revolution did allow advances but the main factors were the repetition and scale of what humans could already do. Then came the advances of precision engineering and manufacturing. Map that to knowledge work and you have a world changing technology but not the silicon valley vision of a digital god. And one which doesn’t necessarily have the moats some see either.

It’s part of why I don’t see the economics of this adding up though. The labs and big companies cannot carry on doing what they doing forever. It’s burning money.

 

[EndME]

So this is the thing that I think is far more likely than some superhuman intelligence. It’s human level intelligence or rather knowledge work as a commodity...Map that to knowledge work and you have a world changing technology but not the silicon valley vision of a digital god.

I see no reason to postulate either of these things when there's an infinite range of possibilities in between them. It doesn't seem to work like a human and there's no reason to think of it as what people would typically think of a god. Things are allowed to be different. And I believe it's clear that it is doing things humans couldn't already do.

It’s part of why I don’t see the economics of this adding up though. The labs and big companies cannot carry on doing what they doing forever. It’s burning money.

It doesn't have to and they don't have to. Probably within a year a five year old will be able to download some weights and do things most humans didn't think were ever possible. Where will that leave those that used to do those things for a living? Do those economies still make sense? I think not, something new will be necessary and the hope is that this should be beneficial to all, but the outcome seems less clear.

 

[mariovitali]

Thanks to everyone for your replies. A family member is in the hospital and the available time at the moment is minimal. I will post to another thread a proposed hypothesis which was identified via AI but -to the best of my knowledge- has a biological basis worth to be considered.

 

[hotblack]

An interesting paper with more work from the Baker lab using AI to help design proteins

De novo design of cysteine proteases

Spoiler: Authors

Choi, Hojae; Bauer, Magnus S.; Coventry, Brian; Venkatesh, Preetham; Chen, Anqi; Kim, Donghyo; Bera, Asim K.; Kang, Alex; Nguyen, Hannah; Sadre, Shayan; Decarreau, Justin; Joyce, Emily; Shankaran, Bhanumathi; Thompson, Tuscan Rock; Greenstein, Gabriel; Didi, Kieran; Schaaf, Lars L.; Gershon, Jacob Merle; Shida, Alexander F.; Lee, Gyu Rie; Hilvert, Donald; Pellock, Samuel J.; Baker, David

Abstract
Despite advances in de novo enzyme design, success has been largely limited to low-energy-barrier model reactions. Amide bonds such as those linking amino acids along the peptide backbone are stable for hundreds of years in neutral aqueous solution because of the high energy barrier to hydrolysis.

Here we describe the de novo design of enzymes that use an activated cysteine nucleophile to hydrolyse the polypeptide backbone in a sequence-dependent manner, with a success rate of 13/69 = 19% and rate enhancements over the background reaction (kcat/kuncat) of up to 3 × 107. The designed proteases have folds that are very different from proteases in nature (TM score < 0.50), and six crystal structures are very close to the design models (Cα root mean squared deviation of <1.2 Å), highlighting the capacity for generalization and the accuracy of the design methodology.

Experimental and computational analyses suggest that the remaining gap in activity to the most active native cysteine proteases arises from imperfections in active-site preorganization and substrate positioning. The designed proteases efficiently cleave their targets in mammalian cells, opening the door to a wide range of synthetic biology applications.

Web | DOI | PDF | Nature

Also see the discussion here on Ars.

A protein is a chain of amino acids held together by what are called peptide bonds. On their own, those bonds are extremely stable; they’re estimated to have a half-life of several hundred years. But in Earth’s environment, most proteins don’t last nearly that long. That’s because organisms have evolved a handful of ways to catalyze the break-up of peptide bonds, which have diversified into countless specialized protein-digesting enzymes. Proteases are everywhere and generally make short work of any proteins that find their way into the environment.

Now, researchers have basically short-circuited the diversification process, using AI-based protein design tools to take the catalytic core of two types of protease and replace the entire rest of the protein. The resulting enzymes could be tuned to cut specific proteins in half, and the researchers demonstrated their effectiveness by using them to cut proteins associated with neurological diseases

The potential therapeutic use of this is complicated. Proteins are often difficult to deliver to the site where they’re needed, and can often trigger an immune response. Things get even more challenging if the protease target is inside a cell rather than on the exterior. That would likely require the target cells to make the protease, which would involve delivering the gene encoding it to the cells where it’s needed. So, there are considerable hurdles that would need to be cleared before we could potentially start developing therapies using this technology.

The researchers also highlight how the work shows some limitations of the tools. As used, they successfully designed proteins that matched all the physical constraints they were given: They folded up into structures that brought the amino acids needed to catalyze a reaction into roughly the right location, and latched on to specific target amino acid sequences.

But we don’t, at the moment, have the ability to tell these tools we want any specific catalytic activity

 

[mariovitali]

I wanted to follow up on previous posts and conclude here. As a side note, I mentioned that a family member is in the hospital (since last Thursday) and I am mainly responsible for this person. I am currrenty not sleeping too much and being active all day to keep up with responsibilities. 5 hours of sleep , active until 24:00 and energy levels are holding up. A reminder of my personal miracle on managing my ilness.

I will now describe how I have been using AI :

1) A repository is used which contains genes identified and certain studies (Unfortunately this selection is done by me, I could use some help here)

2) Several AI algorithms attempt to suggest a causal hypothesis based on certain importance weights (e.g.DecodeME receives a high weight, so are studies of PEM with temporal analysis , for example Maureen Hanson Study )

3) After each AI algorithm suggests a hypothesis, all AIs are being presented with other hypotheses and express their opinion, if some erroneous information was found this is also flagged. They then rank their hypothesis across all other hypotheses. A final ranking is being made at the end of the process.

4) A file is used to keep research progress across runs which contains all hypotheses, unresolved questions , erroneous suggestions and critical dependency points (=we need BiP/GRP78 and XBP1 splicing to confirm whether an actual unfolded-protein response is taking place. if this is not happening, then hypotheses #6,#10,#12 are not considered valid anymore). Unfortunately, I cannot show here even parts of this file as it is generated by AI.

I do understand the concerns of information explosion in this forum. @Utsikt I believe made a suggestion about a subforum. Of course, I could always just don't bother and start my own website where I report the whole process and then start asking around hoping to be noticed. Problem sorted (?).

Things have progressed in AI since last year. With proper methodology AI can move our research efforts extremely fast. We can be cautious and we have the control to do it in a pace that we can handle.

 

[hotblack]

(Edit: to avoid any confusion this was a cross post with the above post, I just want to make clear it wasn’t and shouldn’t be seen as a reply or response)

One thing I think is likely is a lot of things will be replaced better with decision or ‘system one’ models. They seem like they could do a lot of the grunt work. So much of knowledge work fits that sort of structured decisions.

And despite the ability to scale or these sorts of models being faster, resource constraints will matter. I lnow I bang on about this but it’s like not like were short of intellectual capacity or potential on this planet. That’s not been the limiting factor. Never has. the limiting factors lie elsewhere. We waste a huge amount of human potential and intelligence every second of every year.

Humans may have problems but I wouldn’t put faith in the perfect morals of the machines just yet either, this is an interesting example and there are countless others thst we cannot simply dismiss and assume ‘it’ll all be fine’.

Jev Meets the Trolley Problem · Zach Landes

I asked Jev and 13 other decision models the Trolley Problem.

zachlandes.com

I see huge possibilities and support the technology as a tool. But the way in which the problems, concerns or limitations are dismissed by some in the tech industry is I think part of why there is so much pushback from others. Scepticism is essential in looking at many of the claims.

 

[hotblack]

I do understand the concerns of information explosion in this forum. @Utsikt I believe made a suggestion about a subforum. Of course, I could always just don't bother and start my own website where I report the whole process and then start asking around hoping to be noticed. Problem sorted (?).

Thanks for explaining more about your process @mariovitali and I’m so sorry to hear of the problems you and your family member are having.

For others reading this some of the discussion here is I think a continuation of this thread where @mariovitali is asking about posting of AI generated hypothesis and the discussion continues on the challenges that poses for the forum

 

[EndME]

I wanted to follow up on previous posts and conclude here. As a side note, I mentioned that a family member is in the hospital (since last Thursday) and I am mainly responsible for this person. I am currrenty not sleeping too much and being active all day to keep up with responsibilities. 5 hours of sleep , active until 24:00 and energy levels are holding up. A reminder of my personal miracle on managing my ilness.

I will now describe how I have been using AI :

1) A repository is used which contains genes identified and certain studies (Unfortunately this selection is done by me, I could use some help here)

2) Several AI algorithms attempt to suggest a causal hypothesis based on certain importance weights (e.g.DecodeME receives a high weight, so are studies of PEM with temporal analysis , for example Maureen Hanson Study )

3) After each AI algorithm suggests a hypothesis, all AIs are being presented with other hypotheses and express their opinion, if some erroneous information was found this is also flagged. They then rank their hypothesis across all other hypotheses. A final ranking is being made at the end of the process.

4) A file is used to keep research progress across runs which contains all hypotheses, unresolved questions , erroneous suggestions and critical dependency points (=we need BiP/GRP78 and XBP1 splicing to confirm whether an actual unfolded-protein response is taking place. if this is not happening, then hypotheses #6,#10,#12 are not considered valid anymore). Unfortunately, I cannot show here even parts of this file as it is generated by AI.

I do understand the concerns of information explosion in this forum. @Utsikt I believe made a suggestion about a subforum. Of course, I could always just don't bother and start my own website where I report the whole process and then start asking around hoping to be noticed. Problem sorted (?).

Things have progressed in AI since last year. With proper methodology AI can move our research efforts extremely fast. We can be cautious and we have the control to do it in a pace that we can handle.

Thanks for sharing. Could you please declare which versions of which Algorithms you're using since this is currently the essential part?
I also wonder about the secrecy, as using such tools is now standard practice. I'm aware that in some sense you were way ahead of the curve and had a business model to protect, but is that still the case? I think things would be a lot easier if more things were out in the open.

 

[hotblack]

I will now describe how I have been using AI :

I think it makes a lot of sense to have LLMs digest a big dataset the data in this way and look for patterns or links. One problem I also see is how they can draw things together in ways influenced by the language in that data. We see that in models and its easy to see links when they’re none there. And we come back to verification, a point raised by many. The human filtering of inputs helps but doesn’t eliminate issues and the adversarial or committee agentic approach also helps but doesn’t eliminate IMHO. I get why people are sceptical and tbh scepticism is important here.

That said I can see this approach could be used in coming up with a selection of hypothesis that some people could discuss and interrogate and if they think are worthy then explaining in their own words and getting wider thoughts on. I see this more as a brainstorming and learning tool than an LLM giving us an answer, which is IMHO (and I understand others disagree) not feasible now or in the near future for this sort of thing. Or rather, these tools can easily give us an answer. They can give us many. Which goes back to why we have the rules we do on the forum.

Thanks for sharing. Could you please declare which versions of which Algorithms you're using since this is currently the essential part?

That’s a really interesting idea. An open repository would be a great way to allow people to understand the process as well as the outputs. Both problems solved with one solution perhaps? People could see the hypothesis, the prompts, maybe even the decision making processes of the agents. Other agents could also be pointed at the repository to build upon or challenge it. Human discussion could even take place in the repository which could be used as subsequent input and refinements to hypothesis if relevant. There are a lot of possibilities there. And a git repo may be more suited to human and LLM collaboration than a forum here.

 

[Eleanor]

AMIE's conversations were rated favourably in 87–100% of cases (17 criteria) by clinical evaluators, and 48–96% (16 criteria) by patients.

That's quite a big difference.

 

[mariovitali]

Thanks for sharing. Could you please declare which versions of which Algorithms you're using since this is currently the essential part?
I also wonder about the secrecy, as using such tools is now standard practice. I'm aware that in some sense you were way ahead of the curve and had a business model to protect, but is that still the case? I think things would be a lot easier if more things were out in the open.

At the moment I am using ChatGPT GPT-6, Claude Opus 5 and I am also using POE for other LLMs like Qwen, Google Pro, DeepReasoning etc.

I have not concluded if this will be a business yet, this is a fair question and -hopefully- a fair answer from my part. I do own a patent which I immediately discussed here years ago.

That’s a really interesting idea. An open repository would be a great way to allow people to understand the process as well as the outputs. Both problems solved with one solution perhaps? People could see the hypothesis, the prompts, maybe even the decision making processes of the agents. Other agents could also be pointed at the repository to build upon or challenge it. Human discussion could even take place in the repository which could be used as subsequent input and refinements to hypothesis if relevant. There are a lot of possibilities there. And a git repo may be more suited to human and LLM collaboration than a forum here.

Yes it is an interesting one, I agree. At the moment I have not decided whether I can expose the whole methodology since it can be used to other verticals and thus easily copied but I can provide prompts, reasoning, hypotheses being put forward. I think it's a fair "deal" and it does not pose any hurdles in advancing ME/CFS research. What is more important is the hypotheses being able to be presented and also any critique set forth by AI for any research proposal or hypotheses by other researchers.

If anyone has an objection that witholding the methodology introduces hurdles to medical research I am open -as always- to discuss. I do not wish to introduce any hurdles. The biggest hurdle I see at the moment is that these hypotheses cannot be posted and thus cricised or otherwise be evaluated here.