This summer, due to an unlikely combination of factors, I ended up in the emergency room of a small-town hospital. It was a Friday, in the middle of August, at midnight.
About four and a half hours went by before I was admitted into a room where I spent the rest of the weekend, but that’s another story.
What’s more interesting is where those four and a half hours went.
After the triage, I had one consultation with the doctor, a blood test, a second shorter consultation and an imaging test. A CT scan, more specifically.
All those took about 2 hours total but I would have to still wait 2+ hours more before getting the diagnosis (which wasn’t really definitive either).
That was the time it took the radiologist to analyze the CT scan imagery, write up the report and communicate her conclusions to the doctor. More than 150 minutes.
I assume this had to be done while in between tests for other patients and while processing the backlog of previous reports to be written up.
In my view, this was one of the most crystal-clear cases of lack of AI diffusion I have experienced lately.
Specific AI models and algorithms have proven to be very accurate in imaging diagnosis. Some studies estimate the improvement in accuracy to be in the range of 12% to 29%, while screen-reading workloads fall by about 44%.
Besides the imagery analysis and diagnosis, what is also very clear is that general-purpose LLMs are extremely good and speedy at writing technical, medical prose.
This has caveats. Standalone AI loses in the real world in some cases. It still tends to produce more false positives than humans and does a bit worse on general reads (AI wins on “is there cancer X?”, not on a whole-patient interpretation). But this is improving by the month. Not just because of increased model accuracy but because of the integration of different modalities and tools.
Specialist AI tools now match or beat the average radiologist at specific, well-defined detection tasks. But the strongest real-world results come from AI plus radiologists, not AI alone.
In other words, what is not really debatable is that the combination of AI plus radiologist provides way better results than radiologists alone.
This combo has enormous potential to improve diagnostics, reduce workloads, shrink the time to provide a treatment to patients, all while making imaging diagnosis way cheaper.
In turn, this would enable a dramatic increase in the number of tests to be done which, in another practical instance of Jevons paradox, would create even more jobs for radiologists (a phenomenon already taking place which we might as well call the Jensen’s paradox).
The reality, however, is you don’t really see any hospital or medical group anywhere using this. Or even discussing it.
I am not sure if a combination of generalized lack of awareness within the medical sector, bandwidth issues or perceived incentives (if the Jensen’s paradox above is not understood, professionals might introduce resistance due to fear of losing one’s job) plays a role.
Even if they do, I think the main reason is we may be facing an acute case of the bitter lesson. Because we humans just know better.
In his famous 2019 essay, Rich Sutton concluded that - in the field of AI research - general methods that scale with computing power have consistently beaten methods that build upon human knowledge.
While researchers kept trying to encode HOW, under their views and experience, a problem SHOULD be solved - from chess to visual design - because, you know, they know better, the brute-force approach of letting compute and data do the learning overtook it.
Even when faced with the reality of a technology having the potential to improve your work (by double digits) and allowing you to do more with less (more than 40% less), embracing it takes years.
You have invested years in mastering a skill. Getting into a profession that takes effort and focus to master. And then, you have tech being able to do parts of it better, mostly because of a brute-force approach.
It is in fact bitter.
But resisting is futile. Optimism about the future and an adaptive mind will serve us all better.