The 5 Levels of AI Automation in Healthcare Delivery
It's often hard to understand what we're talking about when discussing "AI in Healthcare." This framework lays out the 5 distinct levels of automation, and a roadmap for implementation
When you hear “Healthcare AI Solution,” do you imagine an automated way to generate billing codes, a holographic doctor who will diagnose and treat, or something in between? The role of artificial intelligence in healthcare delivery is so new that we don’t yet have a common ontology for the different ways that machine learning can impact care, and the associated risks, benefits, opportunities and challenges. Fundamentally, all of these solutions seek to gradually and progressively transfer workload and responsibility from humans to machines, in an area where billions of lives and trillions of dollars are at stake, and the consequences of a single error can be catastrophic.
Learnings from the automotive industry
Fortunately, healthcare is not the first or only industry to face these questions. There are many parallels with the development self-driving cars - where a similar set of mundane through life-and-death decisions are being handed off to a machine. To organize this evolution across myriad companies and technologies, the Society of Automotive Engineers have developed a Level 0-5 framework, where Level 0 is the 1984 Dodge Ram Van I learned to drive on, and Level 5 is a car with no steering wheel. Today in 2023, most modern cars have significant Level 2 features, some companies are rolling out Level 3 features (especially Tesla), and Level 4 is being tested on a very limited basis.
The technical specifications of this hierarchy define different user expectations and performance characteristics at the different levels, helping developers of these systems know when they have what it takes to operate at a certain level. The higher the level, the higher the risks and benefits, and the greater the amount of testing and validation required to implement.
The 5 Levels of AI Automation in Healthcare
The journey to automation in healthcare can similarly be broken down into 5 levels of generally progressive complexity, risk and impact.
Level 1 technologies are purely back office - no clinicians or decision making is involved
Level 2 technologies seek to streamline and improve clinical processes, making them more efficient and user friendly, but do not seek to impact the actual care delivered. We separate Level 2a tools, which are staff and organization focused, from Level 2b, where information is directed at the patient
Level 3 and Level 4 applications provide clinical decision support - the line between them is drawn by the level of complexity/risk that drives a requirement for FDA approval as Software as a medical device (SAMD)
Level 5 is where the algorithm takes autonomous action (ordering a test or medication) without human approval. This will eventually happen, but for the next decade at least is the realm of science fiction
There are applications currently in use and even more under development in all categories 1-4 for use by both payers and providers, but we have not yet scratched the surface of what is possible in any of the categories.
Key factors - MDM, Clinical Workflow, and Patient Interface
We found that 3 key variables defined the different categories of healthcare AI automation.
Impact on Medical Decision Making (MDM): This variable defines whether a system is trying to impact the actual course of clinical care (Levels 3-5), or only facilitate the process that would happen anyway (Levels 1 and 2). The risk profile for interventions at levels 1 and 2 is lower, meaning that you can go live with a lower bar for testing and validation. Level 3 tools very strongly keep a human in the loop (HITL) for the actual decision making, and the algorithms have substantial explainability, which allow them to be classified as Non-Device Clinical Decision Support (CDS), whereas the more black box and automated solutions in Level 4 require FDA certification for Software as a Medical Device(SAMD), which is a much longer and more rigorous process. Level 5, where an algorithm makes testing or treatment decisions without a HITL, is purely science fiction today. While it will undoubtedly be part of medical care in 100 years, we don’t currently have the legal or ethical frameworks to even properly frame the discussion.
Impact on Clinical Workflow: How much does a solution change the lives of actual caregivers? As I noted in my previous article, what really differentiates GenAI solutions from Predictive AI is the ability to do real work that was otherwise done by humans - and thus free up those humans for other things. Level 1 solutions are exclusively back-end, remote from clinical care - which makes them easier to implement (since you don’t have to change clinician behavior), but also limits their ability to deliver transformational change.
Patient Interface: Is this a staff-only tool, or does it affect the patient experience as well? Navigating America’s Byzantine healthcare system, and then understanding complex medical concepts while feeling ill is a Herculean task for even the most resourced and educated patients. A study that showed that ChatGPT is better than most doctors at empathetically educating patients is sadly unsurprising - good communication takes time, and physicians must constantly balance the needs of the patient in front of them with the needs of the six next patients who are already waiting outside the exam room. Combining Patient Interface with MDM Impact has to be done carefully - without a HITL between algorithm and patient on issues that pertain to actual medical care, the FDA rightly expects a high level of validation and SAMD approval.
Where are we today?
In 2023, most providers, payers, and technology companies are starting simple. They feel that there is enough RoI in streamlining the back-end of healthcare administration (Level 1) and unburdening their clinical staff from administrative (Level 2a) and certain patient education tasks (Level 2b) before trying to affect MDM with tools they don’t yet fully understand. In radiology and pathology, there will continue to be development of Level 4 solutions which will have great impact in their focused areas, within a very well defined scope. Level 3 is currently the most under-developed category, and the one with the greatest potential for not just saving money, but improving the quality of care and health outcomes.
The next set of posts, will go through the levels in detail, including a brief survey of the landscape, starting with Level 1 - the Back Office. Stay tuned!