Researchers Train AI Doctors In Hospital Simulation
Key insights of today’s newsletter:
Researchers from the Institute for AI Research (AIR), Tsinghua University, China trained an AI to run a hospital simulation: Agent Hospital.
In the simulation, AI patients go through the full hospital cycle: from registration to consultation, examination, diagnosis, and more.
The simulation allowed the researchers to train AI doctors, or ‘doctors agents’, to handle tens of thousands of cases within just a few days, achieving state-of-the-art performance.
↓ Dive deeper… or listen to the 1-minute podcast episode:
What if you could run a simulation of a hospital? With patients, nurses, doctors, surgery rooms, and everything. And what if you could then simulate the treatment of a patient’s illness across the hospital cycle: disease onset, triage, registration, consultation, medical examination, diagnosis, medicine dispensary, convalescence, and post-hospital follow-up visit?
That’s what researchers from the Institute for AI Research (AIR), Tsinghua University, China, did and published about in their paper: Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents.
Let’s rundown the experiment, why it’s so exciting, and what it means for AI in healthcare.
Welcome to Agent Hospital
One of most underrated capabilities of today’s AI models is their ability to roleplay. And I’m not talking about AI companions here, I’m referring to the capacity for AI agents to simulate believable human behavior.
The concept was first trialed by researchers from Google and Stanford in a now iconic research paper. They demonstrated how twenty-five generative agents interacted with each other in a virtual town called Smallville. They would wake up, cook breakfast, and head to work; they would form opinions, notice each other, and initiate conversations. Over time these generative agents produced believable individual and emergent group behaviors (i.e. “simulacra”), acting like ordinary people living ordinary lives.
In Agent Hospital, the same concept was applied but this time to a hospital environment.
The researchers defined different roles for different agents, from doctors to radiologists to a receptionist handling hospital registrations. They also created a handful of agents visiting the hospital after developing a disease.
During their stay, these patients would move through the hospital process just like a regular patient would.
While most today’s research is focused on training bigger models with more medical knowledge through pre-training and fine-tuning, the researchers demonstrate that running simulations could be a novel strategy to make medical agents smarter.
The hospital can be seen as a platform for training AIs, or ‘doctor agents’, who can treat a potential infinite number of patients:
“Due to the low cost and high efficiency of doctor agent training, we can enable the agent to easily handle tens of thousands of cases within just a few days, achieving that would take a real-world doctor several years to manage.”
If you ask me, this is nothing short of revolutionary.
Real doctors supported by AI
Imagine this: What if every real hospital had a hospital simulation running, in parallel. This hospital simulation would have access to all the same information the human doctors have access to: patient history, test results, real-time monitoring, etc.
And now let’s say that at any point in time a human doctor is looking for a second opinion; all they’d have to do is consult the doctor agent, who’s available 24/7 in the hospital simulation, about the patient in question. The human doctor could ask the doctor agent about the next course of action, discuss the pro’s and con’s of different treatments, and more.
This is the future I see for AI in healthcare: not a replacement of capable, caring humans, but a deep contextual support system powered by AI. And quite frankly, this is not science fiction, it’s real and close-to-being-achievable today.
Join the conversation 🗣
Leave a comment or like this article if it resonated with you.
Get in touch 📥
Shoot me an email at jurgen@cdisglobal.com.
I agree that technology can get there over time, even if it is not there today. I am not a doctor so I could be completely wrong. The doctors do a pattern matching and apply if symptom one and symptom two, then disease one and AI will have much more knowledge than a single doctor in most scenarios. In the scenario where there is not a lot of data, a human doctor will have an edge, or in cases where AI does not give importance to a symptom, a human doctor’s experience tells him otherwise. However, we need to address ethical and other questions. I am just listing a few:
How would we handle a medical error situation? Would we still blame the doctor or AI? Are most doctors not going to start trusting AI over their judgment because they would think AI has more knowledge than them?
A great example comes from Kasparov playing against Deep Blue:
“He later said he was again riled by a move the computer made that was so surprising, so un-machine-like, that he was sure the IBM team had cheated. What it may have been, in fact, was a glitch in Deep Blue’s programming: Faced with too many options and no clear preference, the computer chose a move at random. According to Wired, the move that threw Kasparov off his game and changed the momentum of the match was not a feature, but a bug.”