Notes from "The AI Revolution in Medicine"
“The AI Revolution in Medicine” by Peter Lee, Isaac Cohane , et al. offers a very accessible glimpse into how AI could be used in Healthcare. If you are short on time, Pi provided a good summary. I liked Co-Pilot’s summary as well. Interestingly Opus failed to produce one. That said, reading the book in its entirety allowed a better appreciation of the use cases that are at various stages of being realised.
My takeaways from the book were:
Even experts at the cutting edge, like the authors, didn’t anticipate the potential capabilities that have emerged in GPT-4.
It’s going to be hard to predict the scope, timing of the impact from these new AI techniques without any accuracy as the improvements are faster than Moore’s law and are probably exponential.
This will have far reaching impacts in a decade. On the positive side:
People would have access to a reasonably reliable answer tailored to their time critical medical questions, while they wait for a health care physician. This would be a cheap, life saving tool in countries with poor health infrastructure.
Even when doctors are present, the asymmetry of information in the interaction between patients and doctors in countries like India would decrease significantly. People would be able to sense-check their questions or the doctor’s response.
Western healthcare systems would have much greater throughput by cutting out paperwork which apparently accounts for more than 50% of their time:
Systems would record the conversation between doctors and patients, transcribe them in expected formats
The conversation would be indexed for accurate retrospective
The AI system would suggest various inferences, fairly accurately, that the doctor would edit before authorising
diagnosis
recommendation
insurance codes
correspondence to the patient
The above benefits are likely just the low hanging fruits and are likely in various stages of trial. The potential in the field of medicine seems even broader.
There are some fun challenges/implications outside of Tech/Medical sciences linked to the changes above:
Who bears the liability for the inaccuracy of various components?
What level of inaccuracy from these systems are we willing to tolerate - what are the alternatives?
What aspects of our health data are we willing to (and obligated to) share?
How do we secure, anonymise the data within the expected constraints?
What are the risks of relying on these systems?
I hope to explore these further from the works of experts like Mustafa Suleyman, Eric Topol, etc.