There has been a decades-long evolution towards creating an effective clinical diagnosis and treatment AI system, essentially automating the mental part of what doctors do. A solid basis for the content of the system has already been built in the form of medical text books, procedures, published databases, studies and clinical standards such as HEDIS.
The major elements of a fully automated system have been built and put into practice in a variety of medical practices. When a comprehensive system will be built and deployed is impossible to predict. No fundamentally new tech needs to be invented for this to be created; no “break-throughs” in AI! It “just” needs to be applied.
While having an AI-driven medical diagnosis and treatment system would be amazing, much more important than the AI aspect of it would be the fact that it would be data-driven instead of human-created-policy-driven. This means that the system would, over time, determine what actually works based on the data and results, rather than what human “experts” and self-interested institutions say works. In other words, it would support true evidence-based medicine, replacing the too-often corrupt practice of studies published in medical journals. This is a huge subject.
What do doctors do?
They start with the patient’s complaint, why they’re seeking help.
They then get from the patient and/or medical records a time sequence of conditions (like a cough), tests, observations, events (like falling down), related personal things (age, heritage), and finally diagnoses, treatments and outcomes.
Based on this, they make further observations, tests and measurements. The tests may involve other people and equipment, for example a CAT scan. Depending on the expense and trouble of the test and the chances it will affect the outcome, further tests may be performed.
The result is that the doctor recommends and/or performs treatments that will resolve the issue. The treatments can include drugs and operations. The results of all of this are stored in the patient’s EMR, partly coded data and partly written clinical notes.
In order to do the above, doctors receive a great deal of training, both general and clinical. While in practice, they are guided by their knowledge and experience, and also by clinical guidelines and protocols, which evolve over time.
Doctors are limited by a couple of things. First, missing information: they may not have access to and probably don’t have time to read all the patient’s medical history. Second, missing knowledge: there is a huge and ever-growing body of medical knowledge and treatments. It’s amazing that doctors have as much of this in their heads as they do, and not surprising that they sometimes forget or haven’t had time to read and absorb information that is new to them.
Is all the technology required really available?
The pattern of an innovation being proven and waiting sometimes for decades has been demonstrated many times. For example, an algorithm applied in production more than 50 years ago (!) for optimizing oil refinery operations has only recently been applied to optimizing some aspects of health care scheduling. Here’s a detailed example.
No new math or fancy algorithms are needed. The fancy new AI LLM’s (large language models) that are getting attention these days don’t apply to this problem. The vast majority of the effort is in centralizing, codifying and standardizing data that is entered into medical EMR’s, which has already been done and is being refined. Even the tricky work of extracting value from doctor-written clinical notes is largely automated. Large databases of this kind are in use today by pharma companies to help them discover and refine targets for drugs.
The path to automation
The word “computer” was originally applied to people, mostly women, who spent hours and days bent over desks, often with calculators, computing the result of various mathematical formulas. For example:
Barbara “Barby” Canright joined California’s Jet Propulsion Laboratory in 1939. As the first female “human computer,” her job was to calculate anything from how many rockets were needed to make a plane airborne to what kind of rocket propellants were needed to propel a spacecraft. These calculations were done by hand, with pencil and graph paper, often taking more than a week to complete and filling up six to eight notebooks with data and formulas.
While not as precise, doctors are also human computers, in the sense that they confront a new case (problem), get inputs from the patient and the database of the patient’s history, make observations (like calling a data-gathering subroutine), search their memory for a standard to see what to do next (if X and Y, then do a blood test to see if Z). Depending on the results of that test, there may be further branches (if-then-else) to see what other tests and procedures may be required. Finally you reach a diagnosis and a treatment plan. The results of everything including the diagnosis and plan are recorded in the EMR for the patient to form the basis of future medical interactions.
All of these things are in medical text books, treatment protocols, check lists, medical databases and academic papers. They are all pounded into doctors’ heads by clinical training and apprenticeships. Doctors are expected to remember everything.
The path to automation isn’t fancy. It basically amounts to getting a computer to do what a doctor does: interacting with patient (taking input and providing information), organizing and enhancing the records about the patient, standardizing and digitizing all the existing protocols, and creating digital channels to orders for tests, procedures and drugs. Most of which are already a feature of EMR’s.
Most of the elements of this automation are already in place! WebMD.com, for example, has a huge amount of information about symptoms, diseases and treatments online. It’s medically reviewed, and organized for access by patients. Major hospital systems have similar websites. The websites are just the visible part of the iceberg, with vast underpinnings.
The most obvious missing elements is the ability to request tests and procedures – for that you have to go to a human. But the ability to input requests for such things is already a feature of the EMR’s used by most doctors. Making the connection from the EMR to software instead of a human is a minor task.
Automating doctor decision-making is the heart of the job. It’s essential that this be done using an editable, extensible decision tree. This can be enhanced with probabilities and ever-increasing amounts of personalization. This should not be created by training of any kind; it must be human editable and fully transparent, so that you always can know exactly how and on what basis every decision was made.
Among the biggest missing elements are things that doctors learn during their clinical training and personalization.
Once all these elements are put together and working, you would enter a parallel production phase, in which the computer would get the same inputs a human doctor would and propose what to do next. This would be recorded and compared to what the human doctor did in classic champion/challenger fashion. The system wouldn’t have to be 100% complete to be put into live operation, so long as a good system for bailing out of the computer and shifting to a human doctor was in place. But since such a large number of patient visits are routine, the computer is likely to be able to handle a large fraction of cases from early on.
There is a huge amount more detail in the building of such a system. However, surprisingly little needs to be “invented” to make it work, given that large elements are already built and in production in limited ways.
Related posts
Doctors too often get the wrong answer. This is the kind of thing that makes some people hope that automation could do a better job:
https://www.blackliszt.com/2016/12/what-can-cats-teach-us-about-healthcare.html
Massive spending has gone into "cognitive computing" and healthcare, with nothing to show for it.
https://www.blackliszt.com/2015/07/cognitive-computing-and-healthcare.html
You don’t need AI or cognitive computing to discover or promulgate the new discoveries that humans make.
https://www.blackliszt.com/2015/08/human-implemented-cognitive-computing-healthcare.html
Health systems have trouble just making computers work. When they try to do something "fancy," the results are usually poor. But there are promising exceptions.
Healthcare systems spend huge amounts of money on things related to AI, but they don't know what they're doing and neglect to spend on simple things that could make an immediate difference.
https://www.blackliszt.com/2016/09/healthcare-innovation-from-washing-hands-to-ai.html
Avoiding error is hugely important.
https://www.blackliszt.com/2017/06/how-to-avoid-cutting-off-breasts-by-mistake.html
A major lesson from the above posts is this: while AI can certainly automate what doctors do, having the usual major corporations and medical systems be in charge of the effort guarantees failure -- which billions in wasted spending to date demonstrates.
The benefits of medical automation
The potential benefits of automation are huge.
Cost of medical care: As medical workers are replaced by software, costs will go down. Not just salaries, but also office space, etc.
Medical care waiting times: The software doctor is available 24 by 7, no scheduling required.
Accuracy of care: Medical people can’t be as consistent or up to date as data-driven software. Elaborate measures such as HEDIS for judging medical care after the fact will be applied as the care is delivered, assuring its accuracy.
Transformation of care: Dramatically better health and lower costs will result once the system is in place and real-world evidence from it supplements, personalizes and replaces existing care practices.
Automation of medical care isn’t without problems. The institutional obstacles are huge. Mountains of regulations and standard practices would have to be changed, with entrenched forces fighting every step of the way. The people whose jobs are threatened will resist. A large number of patients value interacting with a human doctor. Corporate forces will fight to have their interests supported in the rules and data of the automation. There will have to be a way to provide alternatives and avoid centralized government control, which will be a major struggle, and a danger I fear.
Conclusion
Automation of medical care has been underway for decades. All the technical elements to enable it are available. The benefits of automation are large, but so are the obstacles to implementation. Centralized control of medical diagnosis and practice is already strong, and automation would make it stronger and less visible. The path forward is likely to remain slow. While there are substantial potential benefits in terms of cost reduction, better time and accuracy, the largest potential benefits of huge cost reduction and improved patient health are threatened by a version of the centralized control embedded in the current partly-mechanized system being translated to the automated one.