The transformer architecture has revolutionized AI, enabling systems to capture complex non-linear relationships in vast datasets. In medicine, this has led to remarkable capabilities:
ℹ️Current Applications
We clinicians will, or already are, using AI tools at work. It’s crucial that we, as a field, speak the same language as those implementing these tools. This is to ensure patient safety (Epic’s Sepsis cautionary tale) and to use the tools properly. They are quite good, and we should make the most of them.
A crucial distinction often missed is that an AI model itself is not a product. Take OpenAI as an example - while they excel at building powerful models, their success with ChatGPT comes from transforming that model into a helpful assistant. As highlighted in this brilliant Stanford talk, considering the specific context and software surrounding the model allows us to be imaginative and practical.
Consider clinical decision support in radiology. While companies focus on creating high-performance diagnostic models, the implementation pathway remains unclear. There is practical use in screening and translating reports for patient understanding, but clinical practice implementation remains murky.
Currently, using the model, the main product being created is one that generates imaging reports. Here are some options:
Without sufficient thought to human-computer interaction, it’s looking pretty bleak.
Options 1, 2 and likely 3 cause time-poor and stressed out radiologists. Option 1’s ‘helpful’ reporter product is like a genius who sometimes gets the hardest question right and sometimes the easiest question wrong. In a healthcare setting, there is limited value - more time will be spent on all discordant cases (which may not even result in better clinical performance). Option 2 is option 1 in disguise - you risk over-reliance or ignoring useful outputs. Option 3 is more useful; it sets clear boundaries on the human-AI relationship. By only making the AI visible in discordant cases, it may serve as a good tool to ‘triage’ scans up the chain of experience. However, you run into the same ‘Who is right?’ dilemma.
Financially, only option 4 makes sense to radiology practices and hospitals. Ide & Talamas describe this as an autonomous agent replacing routine work, displacing humans to more specialised problem-solving. If this leads to better patient outcomes, we must choose this option. However, we also need to face significant restructuring of training programs and retrain displaced early-career specialists.
Our limited options stem from several unfortunate assumptions/starting points:
Reading medical imaging itself is a process. Why can’t we have asked questions like:
Outside of a resource-poor setting, there is little unmet clinical need for an autonomous radiologist agent. The explosion of AI, the abundance of radiology reports and the monetary value in creating a high-quality autonomous agent all culminates in these foundation models that can perform exceptionally well.
However, given its training with human-labelled reports and diagnoses, I question if we can truly grow in medicine with these types of models. Can we get closer to ‘perfect medicine’ by having models that talk and breathe our biases?
Here is a direction I think would be more fruitful, we already have high-quality intelligent staff, why can’t we empower them to perform efficiently and improve to be their best? All of those 6 questions I’ve posed that aim to directly augment a radiologist’s work are tractable now. Note that they are useful products, not necessarily new models.
Unsupervised data-driven approaches can teach us so much about biomedicine - medicine will look incredibly different in the upcoming decades. We need nimble well-supported staff, with both autonomous AI and better non-autonomous copilots to maximise their clinical impact.
We’ll explore non-autonomous copilots and autonomous AI in more detail here including specifics of how we can think about human-AI interaction.