Doctors and government departments responsible for building out our healthcare system can’t hope to change at the speed of AI, so what can they do?
In discussion last week with Marc Paradis, a New York-based healthcare data and AI strategist, who has spent most of his life working as an operator inside big US healthcare organisations on both the provider and payor sides, I posited that while the role of doctors was about to be turned on its head, the future for doctors was bright.
The smart ones, I said, would go upstream in the care cycle and do much more interesting interpretive work based on engaging with where AI is taking medicine, both on the provider side and the patient side.
Mr Paradis wasn’t so sure.
“I don’t know that I would tell a 15-year-old to pursue a career in medicine today,” he said.
This article originally ran on TMR’s sister site, Health Services Daily. TMR readers can sign up for a discounted subscription.
He wasn’t saying that being a doctor in the future would be a terrible job. He was just pointing out medicine is entering a decade-plus transition in which human-based training, governance, and workflows will lag the reality of AI-driven care.
Pre-AI training in this vortex of massive disruption and change will leave the next generation of clinicians neither educated nor trained properly for either world. The question is not whether clinicians remain central—they do—but whether institutions redesign quickly enough to preserve human judgment where it matters most.
Best wait a few more years for things to settle down. How many years? Somewhere between 10 and 20 says Mr Paradis.
What the new AI medical world might end up looking like for doctors Mr Paradis isn’t entirely sure. There is too much change coming with too many variables to make predictions with any real confidence.
But he is clear about some things: what the key changes that doctors and system builders are going to have to deal with in the near term, and that, as hard as it may seem, they need to start thinking and planning about how to manage these changes now.
“This is absolutely existential … that you set aside some amount of your revenue and start up an R&D unit now. If you do not do this, you will be out of business in 10 years, I guarantee you.”
Mr Paradis doesn’t have a full grasp of the Australian healthcare system set-up yet but when I explain the basics he suggests that organisations like the Department of Health, Disability and Ageing, the Australian Digital Health Agency, the RACGP and all the doctor colleges, should all have some sort of skunk works in play now to look at how the obvious changes coming their way could be managed.
He doesn’t expect that any of these groups can significantly change their plans and protocols immediately given just how complex things like clinical governance and digital health infrastructure planning are.
But he does think that anyone not thinking about AI now is sowing the seeds of their own demise and potentially of those they are supposed to look after.
What are those key issues Mr Paradis believes are now in play for doctors and policymakers?
He’s got a pretty long list, but as a starting point he cites how AI is flipping the traditional medical model of diagnosis, how that will radically change how doctors work, and, how crucial it is going to be to rethink how we manage clinical governance.
Related
The old frequentist vs the new Bayesian model of diagnosis (no, seriously, read on)
For those who are mathematically challenged by this sub-heading, a quick translation might be that AI is rapidly starting to enable diagnosis to move from being episodic, largely overseen by doctors, to continuous, in the hands of both the patient and doctor.
Mr Paradis describes it as “moving to a learning health system model, where every episode of care, every care interaction adds to the overall knowledge for all patients and all clinicians”.
“AI front ends and AI front doors will meet patients where they are living their life, where 99.9% of health occurs, and this is where, increasingly, diagnosis, along with health, wellness, maintenance and prevention, will be carved out,” said Mr Paradis.
“We have to start thinking of diagnosis differently.”
Can you imagine what a typically over-worked GP today might make of Mr Paradis’ thinking on the future of the core capability of GPs: evidence-based diagnosis?
Not a lot, you’d have to think.
Futuristic and muddled speculation? Maybe not.
All the technology elements are in place for what Mr Paradis is describing to start happening today. Indeed, he can quote a few instances of where it is happening now.
For a GP today diagnosis happens when a patient feels bad enough to book an appointment and maybe they even travel to a bricks-and-mortar clinic, after which they might get tests. This process is episodic, usually late in the disease state process and always driven by symptoms.
In the model Paradis is describing:
- Diagnosis increasingly comes from continuous data streams, not just episodic visits where data – at least in Australia – is stored in a siloed EHR not accessible to either the patient or other providers.
- A patient is not only feeding their health AI assistant their continuous data but data from things like their My Health Record and from all their episodic doctor visits.
- Where appropriate, the data from one patient becomes available as pre-existing learning for many patients. The system is iterating, learning and building.
- AI builds up data where what a GP might have seen once as noise – say, in an EKG – AI sees a signal that is important to a patient’s ongoing health, maybe an early warning signal on their heart condition, which a GP can’t detect looking at a chart.
There’s a huge change here for doctors. Rather than being disease-condition detectives starting from scratch, they start becoming interpreters of all the data at hand, in partnership with their patient, and AI.
A simple way to think of such a change for a doctor might be from “is this the right textbook diagnosis” to “what is the right decision for my patient given what we both understand at this point of time”.
Clinical governance needs to be reinvented rapidly for AI
Mr Paradis, like most professionals working in healthcare, sees clinical governance as a baseline for patient safety that must always in some form be understood clearly and in place.
But he says the clinical governance of an AI medical paradigm is worlds apart from what we largely have in place now.
Not that what we have in place now is bad – just that as AI changes the nature of diagnosis and patient management at speed and scale in an iterative manner so what we have in place now could end up being a major anchor for system development and patient experience.
“We’ve tied ourselves in knots in the West with the ‘governance of no’ and regulation solely for the elimination of risk. ‘First do no harm’ does not mean ‘first do nothing’,” he said.
Mr Paradis says every provider organisation and government agency should be thinking hard about this problem now: current clinical governance has been built to manage fixed, slow-changing interventions and a world of episodic care.
That won’t work in AI medicine which is dynamic, data driven and continuously learning.
Our current clinical governance more or less demands we prove something thoroughly once and leave the setting alone until someone manages the next big evidence-based change. That’s just not going to work for AI, which is living and learning day by day.
How might you even begin to think about such a complex transition of something so fundamental to patient safety?
Mr Paradis doesn’t have the answer. No one does yet. But he is saying to system planner doctors, start working on this problem today, not when AI is moving at the speed of silicon and we find ourselves in a total mess between the two worlds.
“You can create meaningful AI governance where the governance actually enables you to deploy safe, effective, efficient, and impactful AI faster,” he said.
“We need ethics, we need regulation, we need guardrails, we need governance 1000% but we need these things to help us change, adapt, and innovate faster, not to slow us down.”
What else?
How long is a piece of string?
Other obvious structural changes that planners need to model according to Mr Paradis include things like education, training and CPD, institutional system design, digital infrastructure planning and workforce planning.
What does training, education and CPD look like in an AI medical paradigm?
How soon should we start addressing this obvious emerging issue for new medical trainees, and in the case of CPD, existing doctors, and how?
If you map out what Mr Paradis is projecting in terms of the diagnosis side of things, then our current CPD system for doctors in Australia may soon be largely redundant. The information will be held in AI medical databases now being built out at speed like Open Evidence or Heidi’s new venture, Heidi Evidence.
Yes, there will be a need for understanding of the basics, but between the concerned patient working their patient-side AI agent and these new fast evolving and constantly learning data bases, the days of doing CPD learning of detailed iterative changes to treatment and diagnosis are probably numbered.
How might the RACGP begin to think about remodelling its current bureaucratic static learning 50-hour CPD regime to optimise how doctors use the new AI agents to do their work?
How will medical schools engage in this change so they don’t graduate students lacking the requisite skills to manage data and AI.
And so on.
System design and digital infrastructure
A big theme for Mr Paradis is that “necessity is now rising far faster than capacity”.
He provides a very high-level framework for those wanting to start engaging in the transition by splitting the problem into three terrains of care:
- physical clinical terrain – hospitals, GP clinics, health centres and so on where we do trauma, intervention and life and death;
- the virtual clinical terrain where both patients and professionals will manage a move to longitudinal care management via data and through that more effective diagnosis and triage; and
- the population terrain, which will become AI-native population health data in real time and the level of the individual.
In this new world he says we are already witnessing significant institutional capacity limitations – read, building more hospitals is very clearly not a sustainable fiscal model for the future of healthcare.
In this world we know already for certain we are going to run out of workforce to keep running that model. Fiscal and workforce strain without adaptive change will soon lead to a reduction in flexibility and workforce resilience.
Instability will follow and as a reaction system planners will often resort to defensive “no” type governance to avoid the increasing risk.
That all has a bad ending, hence why Mr Paradis is saying planners need to start thinking much harder now and, at least in isolation, start testing how transition could work.
In all of this process Mr Paradis suggests our attitude to clinical governance is likely going to be the key.
“Acceleration without structure and capability without architecture will lead to small errors propagating at scale with the potential to create harm at scale,” he says.
The flipside is the problem of sticking to our current ridged clinical governance models in which structure prevents learning.
In this world, “progress will remain linear while complexity compounds and the gap between necessity and capacity will grow wide rapidly”.
We are going to need some sort of goldilocks clinical governance recipe for AI and a plan for transition.
Workforce
As things stand today, if you took the annual output of every nursing school in the world, it would not resolve India’s current structural nursing gap.
No western country can reconcile their projected healthcare professional workforce shortages moving towards 2030.
In Australia, by that time we are forecasting a nurse shortage of about 120,000 and a doctor shortage somewhere between 10,000 and 20,000.
And none of this is properly contemplating our rapidly emerging aged care burden. It simply isn’t possible to train our way out of this mismatch, or to resolve it with a few extra percentage points of throughput on existing, carbon-limited models of care.
Says Mr Paradis:
“Continuous, multi-modal data streams will be the foundation to really re-engineer and reimagine healthcare in a way that allows care to scale at the rate of silicon (AI), because we simply can’t scale care fast enough at the rate of carbon (humans).”
AI might have the most impact in countries where the ratio of doctors and nurses to service populations is 1:10,000, 1:100,000 or more, suggests Mr Paradis.
“A 40th percentile AI radiologist is orders of magnitude better than no radiologist at all, right? This can literally bring care to hundreds of millions of people who otherwise would not receive care and who, under the existing system, could never possibly receive care [by human supply alone].”
Apply this logic to countries where the ratios are 1:1000 and how do we reorganise around what our current profile of medical providers looks like?
Mr Paradis says:
“Consider the fact that for any given clinical decision, if AI is able to outperform 50.1% of the clinicians who would otherwise be making that decision, then by definition the AI-mediated care of tomorrow is returning better population health outcomes overall than the human-delivered care of today”.
Redesign is not option
When you describe some very admirable infrastructure plans around digital health and data sharing in Australia to Mr Paradis, he starts shaking his head.
“You’ve got some pretty good structural strengths in Australia as a starting point … some good policy and regulation coherence, intentional governance and solid digital foundations,” he said.
(I suggest the latter is still pretty illusory and he laughs knowingly.)
But he says that if Australia is anything like the rest of the world, most people will still have their head in the sand on the speed and scope of what is actually coming at them with AI.
A big near-term change for the system is going to be the patient-side AI agents, which will rapidly start becoming the patients’ digital front door into the system, at scale.
“Healthcare no longer scales,” he says. “Redesign is not an option now. It’s an obligation.
“We must align the rate at which [AI capability] expands with the rate at which the system can learn and adapt.”
Marc Paradis is principal consultant and founder of New York based SIYOM Consulting LLC, a boutique advisory specialising in data and AI strategy for healthcare and life sciences. If you wish to contact Marc about this story you can reach him at marcdparadis@siyomconsulting.com.
