A version of what most people think of as AI (enhanced random forest decision tree machine learning) can indeed first automate medical diagnosis, and then go on to use the feedback data to personalize and enhance medical diagnosis. The result should be fast, accurate and cost-effective, and should greatly reduce medical costs. However, not only will there be massive resistance, there is an even larger danger with automation.
The training of doctors
Doctors undergo a huge amount of expensive, challenging education. After high school, they have to get through 4 years of college, 4 years of medical school, and a minimum of 3 more years of internship/residency. They have to pass multi-hour tests along the way. By the time they fully enter independent practice, they have huge amounts of knowledge in their heads, along with lots of practical knowledge about diagnosis, treatment and outcomes. And of course they are thoroughly inculcated with a wide variety of medical standards, which they are required to meet in order to remain in good standing.
It doesn’t end there! There are ongoing efforts to organize and systematize this knowledge to make it easier to apply in practice. The are general clinical practice guidelines. There are highly specific flow charts for cases published in journals. There are further modifications of guidelines often called clinical pathways created by local care groups that adjust guidelines for their own practices and standards. None of this replaces the extensive training of doctors – it’s meant to add to what they know and/or refine/correct what they think they know. How can they possibly keep up?
Introducing AI to these amazing doctors
Now they’re in practice in a large medical system and some administrator comes along and tells them some AI program (or whatever) is:
- Available for them to consult if needed, or
- Will look over their shoulder and evaluate everything they do, or
- Somewhere in the middle.
How is this going to go? We already know. The multi-billion dollar flop of IBM’s Watson Health gives us the answer. Here is the post I wrote at the start of the Watson furor ten years ago. I predicted failure. It failed.
So how can AI improve medical diagnosis? Do we need better AI? There is a solution. Variations of it have been proven at scale in other industries.
The core of the issue is this: Watson is fed all the knowledge doctors are given in the form of written language. Watson works with the language. Note that for doctors, textbook training is a small part of the overall education – the bulk of it is clinical! You see, hear, touch and get responses. You get lots of real-life cases and handle them from start to finish. You can only learn so much from descriptions of injuries; there is nothing like seeing them and listening to the person who has them.
“All” that AI does is regurgitate a small part of what doctors learn during their 11 year journey to doctor-hood. Little bits of what’s in AI may fill holes or correct tiny parts of a doctor’s knowledge, but that’s a small gain for a huge disruption and time-sink. No wonder it gets rejected.
The path to success
The key to success is as I described here to start from scratch and copy relevant successes from other domains.
One major multi-domain success pattern of automation and optimization is to use the superior technology to replace the people doing the work manually. Not augment/help/advise; replace. Of course there are true experts guiding the new technology and extending/correcting it as needed.
The pattern showed itself early in the Jacquard loom. In oil refinery optimization, teams of skilled engineers were replaced by optimization software that got better results in the 1960’s. Similar changes were made in retail inventory management and replacement part stocking. Mortgage and other personal loans were performed by skilled bank managers and are now entirely electronic. Everyone involved in the pre-automation versions of those efforts (and many others) believed that the personal element was crucial and impossible to replace. It undoubtedly was crucial when people were doing the job; but the effectiveness of the automation more than made up for whatever the “personal element” was adding.
A case I saw over ten years ago was a major computer and software technology support operation supporting major vendors such as Dell computer. The people answering the phone were taught how to use the computer system, but nothing about the systems they were supporting; the training was less than a week. The operator mostly needed to learn how to take his cues and direction from the software. The user had no way of knowing if what he’s being asked to do or say has been done by many people for years, or is a new instruction just for this unusual situation.
This approach enabled every customer service person to be consistent, completely up-to-date, and even personalized based on information known or gathered about the person needing help. You avoid the painful process of building customer service training materials, training the trainers, getting everyone into classes, only in the end to have inconsistent, incomplete and out-of-date execution of your intentions. Now of course the operators could be replaced by computer voice like Alexa or Siri for self-service.
This case is directly relevant to translating to automated medical diagnosis: instead of delivering up-to-date knowledge to a human expert, the computer system is the ever improving expert, ultimately monitored by a small number of human experts. This post describes in more detail the issue of the computer interface. The knowledge in the computer system is complete, up-to-date, personalized and has all the knowledge both generalists and specialists have.
Current medical practice already has support staff performing things like taking measurements, drawing blood, etc. Visual inspection of the body can be done by camera and analyzed by computer better than humans. Similarly, medical images (CAT scans, MRI, etc.) can already be read by software more accurately than humans, but the medical establishment refuses to adopt it.
The people who are being replaced by automation never welcome being replaced. The greater their training, expertise and status, the more they resist. This is a huge issue. Most automation efforts to date have downplayed those issues, saying that technology will “help” doctors. No it won’t. If it’s done right, it will put most of them out of work, the same way cars and trucks put horses out of work.
What is the nature of the technology that does this? The core of the technology is an extended version of what in Machine Learning is called "random forest." This is effectively a collection of decision trees just like you see in many medical journal papers. The trees need to be extended to incorporate more details about the patient and their medical history than is normally done in medical papers, and also more alternatives with probabilities, costs, risks and benefits.
This infrastructure would be ideally suited to accommodating deeply personalized diagnosis, taking into account the individual's DNA, specific responses and other things. This article describes the approach as a path to dramatic improvement, an alternative to the expensive and impersonal RCT trials that are today's gold standard.
What I've described could be implemented today. No massive computing centers, no LLM's or anything particularly new. Some work would have to be done to add ongoing monitoring of results to provide the basis for modifying and extending the ML models with real-world feedback.
The massive risk of going to automation of diagnosis
The issue that concerns me a great deal is the long-standing, ongoing corruption of medical standards and knowledge by powerful interest groups. Once the practice standards are fully computerized, they can be changed in an instant -- or data-driven change could be blocked. With the continuous closed-loop feedback of patient health that is essential to a computerized system of this kind, the system results will clearly show the expensive, harmful practices that are part of today’s standards for what they are, and should be dropped. Will the powers that be permit this, given the great lengths they have gone to make destructive policies into standard practice? Hundreds of billions of dollars are at stake. No one involved in the massive, ongoing fraud is going to shrug their shoulders and say, "you're right. Sorry."
A system of this kind enables champion/challenger trials to be run at a scale never before attempted at minimal cost. If this were permitted and acted on, we would be able to bid farewell to the vast majority of “wellness” and preventive medicine. This by itself would be a huge contribution to improved health and cost reduction. The powers in charge fought like crazy to maintain the covid "vaccine" fraud, whose value was a small fraction of what's at stake here. I suspect they'd go nuclear over this.
With human doctors at the forefront of healthcare, a significant subset of them fail to follow standard authorized practice. Some of this is just making mistakes. But an important subset is because the doctors who deviate from standard practice know that standard practice is wrong! They know that cholesterol numbers shouldn't be lowered, blood pressure medications should be prescribed rarely instead of regularly, and that much of the rest of preventative medicine is not just a waste of money, it is positively destructive. Some of these doctors speak out and get others to see that they're right. A movement to bring about change starts up. Positive change can result, if only by patients getting educated and refusing treatments that hurt them.
Conclusion
I have addressed this issue recently from a different angle. The key is to realize that we don’t have to invent from scratch the way to automate what doctors do – in fact, doing so is a proven path to failure, as we know from IBM’s Watson. We need to examine carefully the pattern of how automation has taken place in other domains and apply those patterns along with proven-elsewhere techniques to find the sure road to success to doctor automation. But only after we find a way to avoid totalitarian dictatorship imposing corrupted medicine on everyone, with dissenters silenced and punished. In my opinion, the key to achieving this is making medical data and studies open source, following the successful pattern in software. Today's standards for vaccines, for example, are the opposite -- drug makers can't be sued for harm, and their data (such as it is) is kept secret by law