Experimenting With Artificial Intelligence in Health Care

One health care provider looks to bring artificial intelligence to patient care.

Reading Time: 18 min 


Digital Leadership

As organizations rely increasingly on digital technologies, how should they cultivate opportunities and address taking risks in a fast-moving digital market environment?
More in this series
Like what you’re reading?
Join our community

5 free articles per month, $6.95/article thereafter, free newsletter.

$89 $45/Year

Unlimited digital content, quarterly magazine, free newsletter, entire archive.

Sign me up
MIT Sloan Management Review: I saw your title, vice president for population health, and I was curious to know a little more about it. What does it mean to be the vice president of population health, and what is population health?

John Glaser: Let’s start out with what population health is, and then move to why it matters. Basically, population health centers on a group of individuals who share a common health challenge or health situation. They might all be going through cancer, or they all have spouses with dementia, or they’re all 25 and really healthy and you just want to keep them healthy. But they have a common health challenge or characteristic.

We will design a series of strategies and tactics that will keep them healthy and make sure that when they seek health care, it is of high quality and very efficient. Population health is defining and managing whatever strategies and tactics are applied to this group. They are intended to achieve certain health care quality, cost, and person-experience goals.

In a way, population health is like being an educator. An educator would formulate an approach to teach a class of 10-year-olds a particular subject, an approach that achieves certain goals of subject comprehension.

Population health complements individual health care delivery, which occurs when you’re in front of your doctor or nurse. It also complements public health. Public health says, “I want to take steps to ensure the health of the community; for example, that the water’s safe and people are being immunized.” You can imagine the Venn diagrams that illustrate overlaps between population health, individual care, and public health.

The reason it matters — and population health has been around for decades — is that health care is in the early stages of a multidecade, fairly significant shift in the business model of care delivery. The shift is largely driven by the fact that medical care costs so much, and continues to cost so much, consuming an amazing amount of the GDP. It’s also pretty uneven in terms of quality.

How has the business model changed? It’s moving from reactive sick care — you’re sick, you show up, we take care of you — to the proactive management of health, where I’m going to reach out and keep you healthy. It’s moving from fragmented, disconnected care to integrated care across the continuum from a doctor’s office to a hospital to rehab to end-of-life, etc. It’s moving from a fee-for-service model, where I’m going to pay you for volume and activity, to a model in which the doctor and hospital are paid on results — the quality and efficiency of care. This business model shift is being driven by reimbursement change, largely from the federal government, but also state and private-sector purchasers of care.

Partners Healthcare System Inc., for example, is incentivized for doing a great job of managing the health of a population of people with diabetes or a population of children with asthma. They’ll receive a certain amount per person, per child, per year to cover all the costs of the care. If they spend less than the amount, they keep the difference. If they spend more than the amount, they experience a loss. Regardless of financial performance, they must achieve certain quality-of-care and health-status metrics.

To what extent is this shift toward population health made possible by digital technologies? You see this issue now with data analytics and electronic health records.

A range of IT resources is needed to help manage the health of a population. There is a significant need for data about the health of the population to be managed.

If I want to hold you financially and clinically accountable for the care and health of 100,000 people with dementia, you’ll ask, “Well, who are they?” I have to know who they are and be able to characterize them. How far along are they in their dementia? Are they poor? Where do they live? Do they have a spouse or a caregiver who can help them out? Do they speak English? To characterize them, I need all kinds of data. I need electronic health record data, but I also need social determinants of health data. I must gather all this information and then make sure Mrs. Smith in one electronic health record is (or isn’t) the same Mrs. Smith in another electronic health record, because there’s no unifying number here that links people across the board.

Having done that, I ask, “What’s the plan to manage Mrs. Smith’s health and health care? How do I take care of these people?” I can turn to a blue-ribbon panel of doctors to get a core plan for a particular disease, but once I get that plan or that algorithm, I have to lay it over the data, tailor the plan to reflect Mrs. Smith’s needs and capabilities, and look for deviations from the plan. I might say, “I’m going to plan to manage Mrs. Smith’s dementia, but last night her spouse passed away, and suddenly my plan has to be revised, because I was counting on that person to help out and I can’t anymore.” Or “Mrs. Smith had a car accident, and she shouldn’t have been driving, but she was, and was critically hurt.” Suddenly my plan is different.

I must have a plan, and then I must monitor deviations that indicate the plan should be revised. And I should also determine how well the plan is working. What are my measures of quality and efficiency? Am I keeping Mrs. Smith out of assisted living, or should she be in assisted living?

You can’t make these decisions without IT to aggregate the data to characterize someone, to determine what the plan is and whether it needs to be altered, and then to generate a series of metrics that say, I’m doing OK or I’m not doing OK at managing the health of the population.

To help define and manage the plan, we must have care managers. They work for a health system (or an employer or a state Medicaid department), and they’re making sure the plan is working and take steps to remove barriers to the plan. For example, we will have people who are poor or who can’t drive, so we have to get them a ride. Or they live in what we call “a nutritional wasteland.” There’s nothing but liquor stores and convenience stores around them, so if we want them to lose weight, we must get them Meals on Wheels or something along those lines. We should have IT applications for care managers who are committed to making the plan work.

In a lot of cases, to stay healthy, patients must manage themselves. They must manage their weight, make changes to a sedentary lifestyle, and monitor a disease such as congestive heart failure. I need a variety of technologies in the home and through social media to help people stay engaged in managing their health.

Although the changes are driven by reimbursement, you can’t manage the health of a population without a viable set of sophisticated IT.

You talked about this as a multidecade business-model shift. What have been the challenges that the health care industry or hospitals have faced with respect to digitalization, and are those going to be the same challenges they’re going to face going forward?

One of the challenges, if you go back six or seven years ago, would be that the electronic health record adoption was small, but now it’s not. Meaningful Use, a Medicare incentive program, has driven this. There has been a lot of progress in adoption.

The next thing we need is the fluid exchange of data between electronic health records — data interoperability. That’s made some modest progress, but we still have a ways to go before the data about a patient really flows appropriately and efficiently.

The third challenge is gathering and interpreting data about the social components of health. We’re still learning which sets of data really matter in which circumstances. For example, if we want you to get your 10,000 steps a day, what data should we gather to determine the likelihood that you will achieve that goal? For example, how do we motivate people? There are different ways to do it, but I need a set of data to define a motivational construct.

But perhaps the greatest challenge is that we must turn to a series of doctors and say, “You guys have to do a much better job of managing the health of people.” And they respond, “Listen, we actually aren’t always very good at doing that. We know how to take care of them when they’re sick, but we were never trained to manage health, and we’re not well equipped to do that.” To help caregivers manage health, we need care managers, links to social services and resources, new support processes, and a series of IT investments. The biggest challenge is making this transformation.

There’s a big cultural change, a big education change, and a series of process changes that must go on in health care that are difficult. The industry is beginning to go through that. As often happens with industry transformations, these changes are occurring while people and organizations are under great pressure to perform today under the old business model.

And then frankly it’s still not clear that if you do all this population health work, you really will “bend” the cost curve. Medicare has come up with dozens of different population health models and organizational arrangements. Which ones will be the winners? We’re still experimenting with different organizational and reimbursement models. We’ve got a lot of transformation work left to do, and that’s why I think it’ll take decades.

What digital skill sets are most in demand or most lacking in health care organizations?

There are several. But regarding digital ones, we see a whole lot of desire — and we’re not alone — for data scientists. The phrase data scientist is fuzzy and has many meanings. In this case, health care needs people who will help providers when they say, “Tell me what the data means and whether I’m doing a good job or not. Help me think through which data from this long list of social determinants of health I should really be gathering and has significant explanatory power. How do I get this data? How do I deal with uneven data quality? What are the best practices in sharing this data with patients?”

Moreover, who’s the doctor responsible for Mrs. Smith when she sees seven doctors? Often elderly people do see seven (or more) doctors. But which one do I hold accountable for her care?

Many skills are needed, but the biggest one is professionals who can help caregivers work their way through the data and analytics.

Is it fair to attribute the data problems as being around connecting the data to real-world outcomes?

Yes, and how to keep the data clean, ensure that your characterization is accurate, know what the data is telling you, all that.

Is regulation hindering or helping this digital transformation, or both?

Probably both. The biggest payer in the United States is Medicare, and the second biggest is Medicaid. When they come up with a new government reimbursement program, the specifics of that program are defined by regulations. For the government, regulations are the specifics of the business relationship between them (as the purchaser of care) and the provider.

Regulations are necessary and can be helpful, because they bring clarity to the business relationship. If the government wants to financially incent a provider to reduce the number of people who get readmitted to the hospital because they weren’t well managed while they were in the hospital, regulations must be created. When is a readmission necessary or not? How much is the incentive?

Regulations can be overdone. But by and large, they help, because they are the means by which the government — federal and state — becomes explicit about what it wants to have done.

As I look toward the next 10 years, there’s a lot of digital disruption coming — whether it’s artificial intelligence (AI), blockchain, virtual and augmented reality, or autonomous vehicles. Are any of these going to have a bigger impact on health care than others?

AI will have a huge impact, using that term broadly. The term AI is not always that helpful, but the next generation of machine-based intelligence will impact health care in all kinds of ways. For example, it’s the machine listening to the conversation between the doctor and the patient and saying, “Here’s what they’re talking about, and here’s the note that should be generated and the test and medication orders that should be placed.”

There is also what we call the contextually aware electronic health record. The machine says, “I know you’re Dr. Roberts. You’re taking care of Mrs. Smith. She has the three following issues…. Hence, I am going to show you this data and suggest that you take these actions.” Technology is saving you steps and making sure you do the right thing.

It’s the logic that says we’ve got an erratic pattern of medication refills by a patient going on, which indicates deterioration in health management, so we must intervene. Or we’ve got patterns of a patient’s gait at home that indicate potential dementia, so we must do something.

There’s plenty of technology-enabled change coming, but AI will have the biggest impact on health care, population health included.

I read that Cerner Corp. is experimenting with AI. How have those experiments gone? What have you learned?

We have several experiments and efforts. For example, one thing that drives doctors nuts is how much time they spend documenting care. In a couple of clinics, a machine is reading a patient’s electronic health record, watching the doctor’s pattern of use of the electronic health record, listening to the conversation between the doctor and patient, and seeing the physical interaction. From this it generates notes automatically. That’s just pretty neat.

Another example: We are analyzing the cost to treat a particular disease. When I look at the people with that disease, the treatment costs are all over the place. Why is it that some are more expensive than others? Some patients are expensive because they are really sick and they will always be expensive; there is not much I can do. Others will be expensive because we are overtreating; those we can manage better. And because a patient is less expensive does not mean that his or her care is better. Maybe the patient is being shortchanged.

Can I tease apart the reality here and conclude that if you want to reduce cost, go after these five patients (or these five physicians who are overtreating)? Moreover, here’s what to do about these five: This one needs motivational help. This one can’t afford the copayment, so she is not buying her medications.

We use advanced intelligence to help us identify approaches to determine which patients need which types of interventions.

As another example, to characterize and manage a population, you must aggregate data from many sources: dozens and dozens of electronic health records, insurance claims, social determinants, and device data. The data is very messy. We use machine learning techniques to clean up the data; about 95% of our resolutions of data inconsistencies and quality issues are coordinated and done by the machine.

Finally, we use advanced intelligence to monitor treatment effectiveness and efficiency. You can look at the data and pick up patterns that suggest this drug is hurting people, or treatment A is better than treatment B. These kinds of analyses can identify issues and opportunities one to two orders of magnitude faster and less expensively than prior analysis techniques.

It will be a couple of years before many of these efforts become broadly available in health care IT products.

A couple more meaning three to five, or a longer time frame?

Three to five.

How are the doctors who have been involved in the experiment? What has their reaction been? Do they trust it?

They are excited about anything that takes a load off them. It’s hard to be a doctor these days. They’re scientifically trained, they’re very smart, and without stereotyping, they’re pretty much gadget guys, so they often like this stuff.

They also became doctors because they want to take care of patients. Advances that help them be better as caregivers are generally well received.

You mentioned before that this is a big, multidecade shift. When we talk to organizations, they talk about being overwhelmed by having to transform their entire organization. But you’re describing a transformation of an entire industry, which seems much more overwhelming. If culture and process are hard things to change, how do you go about doing that, and does it mean thinking more vertically in some ways? Are organizations like Kaiser Foundation Health Plan Inc., which owns the entire vertical, better positioned to do this, or do we have to figure out how to transform the culture across every layer of the industry?

You’re correct. This is an industry transformation — this is not just doctors and hospitals. You may have seen CVS Health Corp. proposing to buy Aetna Inc. Holy smokes, what is that all about? And there is speculation that Amazon.com Inc. is likely to get into the pharmaceutical benefits management business. Will Apple Inc. significantly expand its health care business? All of a sudden, there are players and industry arrangements that we would not have thought of five years ago.

The business model shift that we talked about is very, very significant and touches all actors in health care. There is a remarkable blurring of traditional organizational boundaries and roles. You see hospitals and health systems establishing health plans. Health plans such as UnitedHealth Group Inc. are becoming providers of medical care. Pharmacies such as CVS are becoming health plans and growing their primary care business.

We don’t know yet what the mature forms of this transformation will look like. We have seen horizontal consolidation, health plans consolidating and health systems being formed. Now we are seeing different forms of vertical integration, such as health systems, setting up health plans. There are several species of vertical integration, with the winning species unclear.

How do you make that change culturally?

I don’t think you can manage this, not really.

I can set certain guardrails that say, “I don’t know how it’s going to play out, and I don’t know what the mature form looks like, and I don’t know how long it’s going to take, but I can make sure people don’t get hurt.” I can also reinforce a business model direction by, for example, managing health in addition to managing sickness, even if I am not sure of the best ways to make this happen. I’m going to try to put up guardrails and strategies that are directionally correct and reduce the likelihood of a calamity.

Providers will need to operate in two business models simultaneously: volume-driven payment and outcomes-driven payment. Many professional societies, such as the American Hospital Association, are helping hospitals figure out how to do that. We may not know how rapidly this change will happen, so we should make sure the participants can survive if the change process is lengthy.

We are experimenting with new models of care, and payment with the winning models is unclear. There is a premium on ensuring that the results of these experiments in models are transparent and rigorously assessed.

It may be that managing the cultural change of an industry is similar to managing the cultural change in a country. You don’t really manage it directly; rather you create conditions that make a particular change more likely than not. If you want water to flow down a valley a certain way, you don’t manage the behavior of individual water molecules. Rather you shape the valley and expect that the water will, more or less, flow the way you want it to flow.

How do you manage that uncertain, chaotic environment when you don’t know what the future’s going to bring?

The best way I can think of is to say, “I’ve got a reasonable range of uncertainty, and there’s a set of viable futures that might occur. I should be doing what is, as best as I can tell, most likely to be relevant in all possible futures.” For example, gathering data about patients to characterize them — I don’t see a future where that’s a bad idea. Show me a future where engaging patients to manage their health is a bad idea, because I don’t see that either. You can do certain things — and, in fact, a lot of the industry is doing them now — where you say, I don’t know how it’s going to play out, but under almost any realistic circumstance, these things will be relevant.

With that approach, you can argue about pace. Should I do it now, or do I have more time? It varies. You see an aggressive movement to this new world in Massachusetts, but it’s a much slower pace in the Dakotas, because they have an unemployment rate of 2%. They’re not about to constrain people’s choices of doctors and hospitals.

The best you can do is to craft the valley and watch the water, pursue a set of activities that are valuable in many scenarios of the future, and create approaches that enable organizations to straddle two business models simultaneously.


Digital Leadership

As organizations rely increasingly on digital technologies, how should they cultivate opportunities and address taking risks in a fast-moving digital market environment?
More in this series

More Like This

Add a comment

You must to post a comment.

First time here? Sign up for a free account: Comment on articles and get access to many more articles.