I asked LinkedIn friends to submit their questions related to AI in healthcare in preparation for an upcoming keynote at this year’s HIMSS in Vegas. I promised to try to answer the questions they submitted. This one is courtesy of my friend, the great Aman Bhandari and it garnered the most “likes.” Here’s his question:

“I have asked for this several times and haven’t received an answer – where will AI increase costs and where will it decrease costs in healthcare? Haven’t seen a cogent response to this, just hype and lots of marketing.”

One reason for the lack of cogent responses is that…wait for it…it depends. I hate that answer as much as you do. The good news is the dependencies are pretty straightforward and easily applied to figure out winners and losers. The two things that one has to know to predict winners and losers:

  1. What “AI” is / does so you can understand the problems it can help solve.
  2. The economics of the stakeholders involved in solving those problems.

First, what does AI let us do better / differently?

Despite the suggestions of some marketing teams, AI is really just math packaged as software. It’s not a sentient being even if they give it human names and it doesn’t magically turn raw data into cures. Usually, the math involved is designed to learn by example. We call that supervised machine learning. When the data involved includes free text (notes, etc) we can turn it into variables that can be fed to the math with an approach called natural language processing. So AI is really, in most cases, just supervised machine learning + natural language processing. If that doesn’t feel complicated or sophisticated enough, here’s a superfluous formula: AI = SML + NLP.

The important part is that this software math helps us:

  • analyze: what’s really going on here?
  • optimize: what should be going on here?
  • prophesize: what’s likely to be going on here?
  • customize: let’s tailor what goes on based on what’s likely to get the best result.

As you can imagine, these capabilities can come in handy in lots of ways in healthcare. And if healthcare was like other industries, the question of how will it save and cost money would be pretty straightforward. Unfortunately, the economics of healthcare are, shall we say, unique. Which leads to the second key to knowing who will win and lose…

The Economics

There is no single $3.2 trillion healthcare industry. It’s more like 20 or so sub-industries. Each of these sub-industries has it’s own economic incentives. To know who the winners and losers will be, we need to understand how each stakeholder makes, saves, or loses money.

If that isn’t nasty enough, the incentives of these sub-industries are often competing. That means that any given application of AI in healthcare is likely to be both cost saving for one stakeholder and costly for another.

For example:

Preventing Unnecessary Admissions. We help insurers prevent admissions by using AI to identify specific opportunities to intervene sooner. For example, using health risk assessments, claims, care management data, call center transcripts, prescription info, etc we can spot elderly people that are likely to fall down, have an adverse drug event, require dementia screening, etc. If our customers can use this information to prevent an avoidable admission, they reduce their costs.

Winners: Insurer, Patients

That same prevented admission is a $1,500 loss to an ambulance company for the ride that never happened. The nearest hospital loses an estimated $11,500 in revenue. The pharmaceutical manufacturer denied of the opportunity to provide new drugs if suggested, $2,500. The pharmacy benefits manager (PBM), $500. The involved physicians, a collective $3,500. If the admission would have followed with a discharge to a specialized nursing facility (SNF) or rehabilitation center which is not at all uncommon within the elderly population, the facility would stand to lose $10,000 – $20,000 per month. If a caregiver is needed….well, you get the point. If you squint your eyes just right you can start to see how US healthcare is 37th in the world in terms of quality but 1st in terms of cost.

Losers: Ambulance company, hospital, pharma, PBM, physicians, SNF, etc.

To know who will win or lose, think of a problem AI can help tackle because of its capabilities. Then consider which stakeholders make or lose money as a result of that problem being solved.

Let’s do another one.

AI in Imaging Interpretation. The potential of AI to automate image interpretation has been getting a lot of attention lately.

Winners: Hospitals and imaging centers that can interpret images faster and more accurately with the use of fewer radiologists. IF and this is a big if, the lower cost is passed on, insurers win by saving money. Second big IF – if the cost savings to insurers results in lower premiums and co-pays, patients win. Faster, more accurate, less expensive interpretation should eventually lead to interpretation available to more people in different parts of the world. That means better care for more people. Of course, there will be all sorts of new policies required and huge questions of liability. That means the lawyers will win too.

Losers: Radiologists. More on that here.

Of course we can pivot our perspective and focus on the stakeholders, then figure out which problems they’re paid to solve that AI can address. Same rules apply.

Hospitals & Clinics

It’s pretty obvious there are thousands of opportunities to apply AI to improve the efficiency of hospitals. Unfortunately, 95% of hospitals are fee for service (paid based on volume and complexity). In other words, they’re not incentivized to use AI to keep people out of hospitals. This is a deadly shame and will cost tens of thousands of unnecessary deaths each year until the incentive structures change. So how will most hospitals begin using AI?

Follow the incentives to understand real cost savings opportunities for AI.

AI for Scheduling Efficiency. AI can be used to identify which patients are likely to no-show or which patients should be routed to which surgical centers based on anticipated complexity. I like this use case because on face value there isn’t downside to matching to more appropriate levels of care & facility. Of course, so much is dependent on how the results of the AI are implemented. Triple booking scheduled clinic slots based on anticipated no shows can be problematic. Just ask United Airlines. We’re doing scheduling optimization now – early stages. Follow or shoot me a note and I’ll share more on this as I learn more.

AI for Billing. Plans and hospitals are already discovering benefit using AI to spot cases where care delivered should be reimbursed at higher rates. This is not a new application of information technology. However, companies are now replacing static, “rules-based” approaches with AI approaches to discover missing HCCDRG, and CPT codes from millions of data points.

Losers: Payors, more often than not the federal government, pay more. However, getting paid appropriately is important. Of course, that doesn’t mean that there isn’t ample opportunity (and temptation) to upcode and we’ve seen some bad actors get in trouble for gaming the system. Expect to see more as it becomes easier to find more cases that represent greater revenues.

AI for Supplies Management. This one is a huge and largely untapped opportunity. Hospitals manage millions of dollars in supplies. Optimization of ordering and management of supplies could lead to millions in cost savings. Even better, unlike much of healthcare business, the calculation of a return on investment for this particular application could be relatively straightforward (it’s really hard to prove a prevented admission).

There are of course exceptions to how AI will help or hurt hospitals because there are different reimbursement models in the hospital world.

Some hospitals are also insurers (Kaiser Permanente, Geisinger, Intermountain, etc.) and stand to realize transformational benefit from integrating the learning made possible by AI in care delivery. Similarly, some of the newer clinics / care offerings provide services at fixed cost such as IoraCareMore, and CityBlock. They too have the potential to discover enormous advantage in using their data to discover efficiencies in the ways they deliver care (i.e., keep people healthy).

Hospitals with Value-Based Contracts. To figure out which parts of a hospital are likely to adopt AI for the purpose of delivering care more effectively, follow the money to their value-based contracts. These contracts are with government and commercial payors and are called accountable care organizations (usually from govt payors), alternative quality contracts (commercial versions), or bundled / episodic payments when focused on the delivery of an end-to-end service (now in play in oncology and for knee and hip surgeries).

Hospitals – to Avoid Penalties. If patients in a hospital are covered by Medicaid or Medicare then the hospital is incentivized to be sure said patients are not readmitted within 30 days for the same or related reasons they were originally admitted for. If a patient is readmitted in this fashion their care will not be paid for by the government. This creates an incentive to use AI to predict / avoid readmissions.

Unfortunately, most that are exploring AI are using it to predict those most likely to be readmitted. If that’s the goal, they’d be better off saving their money and writing a query for older, sicker patients with more admissions in the past 12 months. The real potential is to use all the available data, not just claims, to zero in on those most likely to benefit from specific interventions (see Cyft example above).

Similarly, hospitals are incentivized to avoid hospital acquired infections among Medicaid and Medicare patients as subsequent required care will not be paid for.

See why “it depends”? Let’s hit a couple in the area I’m spending most of my time these days:

Payors / Plans

AI for Matching Individuals to Care Management Interventions.Health plans pay for the care provided. They therefore stand to save money by getting patients the most effective and efficient care possible before unnecessary admissions / events occur. For this, traditional “one size fits all” risk scores will be replaced with the same AI- driven approaches that Google, Facebook, Amazon and others use to match consumers to ads or products – except they’ll be used to save lives. See Cyft example above.

AI for Customer Service. Losing members can be a costly problem – especially for plans that manage chronic / complex populations that invest heavily in stabilizing their members’ health (e.g., dials plans, special needs plans, etc.). Like most of healthcare, customer service is largely reactive – calls come in, customer service reps attempt to resolve issues, if the fail members leave. We’ve seen that AI can be used to proactively identify those likely to leave or to be dissatisfied (and why) making it possible to get out ahead of issues before they lead to lost members.

There are several other plan-specific opportunities. If anyone wants to dive deeper, feel free to post a comment or send me a note: (ldavolio at cyft dot io). For the sake of time let’s hop to another major stakeholder group, the drug dealers.


Opportunities for pharma span from bench to bedside.

AI for Discovery. The human genome project and thousands of subsequent discoveries at the DNA, RNA, and protein levels were made possible by machine learning’s ability to detect patterns across large and often messy data sets. As our understanding of biology deepens thanks to the availability of new data and algorithms capable of learning from it, the drug discovery process is literally being transformed. What was once an entirely hypothesis driven approach where humans posed the questions is shifting toward scientists starting with an outcome and using machine learning to help discover important relationships to that outcome within the data. New discoveries faster = better targets for molecules, more effective drugs, aimed at more of the right people.

Winners: Pharmaceutical companies that learn to harness datasets and AI most effectively.

Losers: Pharmaceutical companies that aren’t able to make the transition to this very different approach to discovery. Expect several of the larger players to flounder with this necessary transition in the short run.

AI for Clinical Trial Enrollment. The greatest hypotheses must still be tested via a randomized controlled trial before they’re ready to be delivered in pill or device form. The clinical trial industry is ripe for disruption and AI will play a pivotal role. Getting enough people to enroll (and stay) in trials is a huge challenge for pharma and the clinical research organizations (CROs) that help execute trials. As our understanding of biology gets more robust the criteria for who can participate in which trials will become more specific, making it even harder to populate trials.

AI can be used to automatically search millions of data points – from social media feeds to electronic health records – to identify potential applicants. Similarly, some organizations are exploring the use of in-home, natural language processing devices (e.g., Alexa, Google Home, etc.) to gather information for trials once people are enrolled.

Winners: Patients as drugs come to market faster at lower cost. Pharma as they shave hundreds of millions off the cost of data collection and populate more trials successfully.

Losers: Pharma and CROs that remain wedded to their current models that rely on a closer network of clinics to recruit and human data collectors.

AI for Phase 4, Post-Market Surveillance. Once a drug or device comes into the market, depending on it’s classification, it may need to be monitored for any unanticipated safety concerns. Today this is an almost entirely manual process that gathers only a fraction of available (and relevant data). This is yet another system that will be fundamentally transformed with the help of AI as it scours many data sets for earlier signs of trouble at a fraction of the cost.

AI to Figure Out Which Doctors to Target for Sales / Early Adoption.Much like Amazon uses data & AI to tailor specific approaches / offerings to its customers so too will pharma figure out who to entice with which offer.

Wrap up

I hope this has been helpful. As AI makes it’s way into healthcare there will be lots of winners and losers in the complex organism that is healthcare. Understanding what AI can enable and the incentives of the stakeholders is the key to knowing how AI will lead to cost savings and for whom.

Finally, there are couple patterns / lessons worth note for anyone hoping to realize cost savings or new revenues via AI. This really deserves a longer post but until then:

First, AI is not a solution. It is a capability that can be packaged into solutions that can increase their effectiveness, often dramatically.

Second, success in any specific application is dependent on integrating context. You can’t install AI and solve problems. Amazon didn’t get where they are by buying AI book selling software from a vendor. Neither did Facebook build their empire by installing a better AI ad seller. What they and many like them have done successfully is integrate learning into nearly all of their processes. Yes, AI is a critical part of their infrastructure. But it’s a means to an end. It’s a tool to help them learn from all of their data as quickly and efficiently as possible. If your healthcare organization is to end up on the winning side of these equations it will be because you’ve successfully integrated learning to realize hundreds if not thousands of incremental improvements – not because you installed a third party AI product. More on that here.

Don’t hesitate to leave a comment, shoot me a note (ldavolio at cyft dot io), or follow along on LinkedIn or Twitter to continue the conversation.