It doesn't take artificial intelligence to tell you that a preventable hospitalization is not good. A hospital is not a bed-and-breakfast. No one says, "hey, for fun, let's get a hospital room overlooking the parking garage this weekend." A preventable hospitalization is by definition one that could have been prevented. Thus, it costs people, insurance companies, businesses, the government, and society considerable time, effort, and resources that could have been diverted to more productive activities. Plus, a hospitalization may expose a patient to potential badness such as hospital food, being separated from friends and family, medication errors, and antibiotic-resistant bacteria. Again human intelligence can tell you all this. Where artificial intelligence may be helpful is in reducing such preventable hospitalizations, and Clover Health is an example of a company aiming to do this.
Hospitalizations that could have been prevented continue to be a major problem. A 2010 report from the Agency for Healthcare Research and Quality (AHRQ) indicated that in 2008 "one out of every ten hospital stays was potentially preventable" and that "potentially preventable acute conditions accounted for 3.9 percent of all hospital stays, and potentially preventable chronic conditions accounted for 6.2 percent." That would translate to a lot of money going down the proverbial toilet. These rates get even higher when you focus on those who are older or economically disadvantaged. A study from the Centers for Medicare and Medicaid Services found that in 2009, among a sample of those eligible for both Medicare and full Medicaid benefits "26 percent of hospitalizations were potentially avoidable."
Why are these potentially preventable hospitalizations occurring? Humans are like cars, except that humans are a lot more complex and rarely have fuzzy dice hanging from them. In order to keep your car out of the mechanic's shop or the scrap heap, you have to regularly check, take care of, and maintain the car. This includes taking preemptive measures such as routinely changing the oil and the tires. Similarly, preventive care is so important for humans, especially those with more complex medical situations like older adults.
Furthermore, recall what Jerry Seinfeld once said on the television show Seinfeld about relationships: "Breaking up is like knocking over a coke machine. You can’t do it in one push, you got to rock it back and forth a few times, and then it goes over." The same can apply to many chronic medical conditions and hospitalizations. A hospitalization does not always occur suddenly with no warning. Rather, often, things slowly get rockier and rockier before a hospitalization then becomes necessary. The trouble is the human body is complex. A pop-up window doesn't just appear on your body saying, "if you don't take these actions, your hospital-free time will expire in T-minus 45 days and counting."
Here's where artificial intelligence, or AI for those who don't have the time to say artificial intelligence, can help. AI is a very broad term that includes any approach where computers or mathematical algorithms can help assess a situation or make a decision.
One example of an AI approach is machine learning. Machine learning is not about you learning how to use machines. It is when a set of computer algorithms try to perform a particular task or generate a specific insight without following explicit instructions. The algorithms will instead look for patterns in data and try to determine what to do based on these patterns. For example, a machine learning algorithm may look for combinations of symptoms, medication usages, and physical exam measurements that are associated with higher likelihood of going to the hospital. The machine essentially learns to know what to find. You could see how machine learning may help identify where things are getting rocky for patients well before they get knocked over and have to go to the hospital.
That's what Clover Health, which was founded in 2014, has been trying to do. Clover Health is a rather unique combination of an insurance-health care services-tech company. As a Medicare Advantage plan, Clover helps to manage the costs of its beneficiary populations. This means that Clover Health has an incentive to keep Medicare Advantage beneficiaries as healthy as possible and out of hospitals. So far, that doesn't sound too much different from traditional insurance companies.
Ah, but here are the differences. Clover Health has been developing and using data and machine learning approaches to proactively figure out where to intervene long before things get so bad that a patient has to go to the hospital. Guided in part by what the data and computers say, they then deploy teams of health care professionals such as doctors, nurses, and behavior health specialists to check on and even visit the homes of Medicare Advantage beneficiaries.
If you think that other insurance companies can do this, Vivek Garipalli, CEO and Co-Founder of Clover Health, has a response: working with machine learning "is not in the DNA of traditional payors. They are financial companies and not tech companies." Andrew Toy, President and Chief Technology Officer for Clover Health, explained how analyzing health care data requires methods that go well beyond those that have been historically used by many insurers: "health care data is overwhelming, often unstructured and unorganized. Traditional data analysis methods can't really help glean the necessary insights."
How about other tech companies then? Garipalli also explained how Clover Health is different from other tech companies, "an insurer is the only entity that can get access to all of the information on patients." This is what he described as "the data advantage" that allows Clover Health to develop and generate more relevant and actionable insights via machine learning. Sophia Chang, MD, MPH, Chief Clinical Informatics Officer for Clover Health added that traditional tech companies often "don't understand the data and issues involved. They may not know how to bridge between what the data says and what actions need to be taken."
Then there is the third leg of the stool that many insurance companies and tech companies don't and can't do: actual directly caring for patients. Chang emphasized the importance of offering not just office-based care but care at the patient's home. It helps to see what is actually happening in the patient's home environment and what may be influencing his or her health, and "it offers members the peace-of-mind of being in their own environment and avoids the discomfort (and risks) of hospital care." Garipalli related how both his "parents are physicians. I grew up with health care as a topic," and how he spent years on the operational side of health care.
"In 2005, I bought a sleep disorder diagnosis center and brought in local doctors," he added. "This made me realize how local health care was and gained empathy for physicians and the challenges that they faced." Subsequently, he "bought a hospital out of bankruptcy and turned it around to profitability." This experience also made him "grow cynical about the business of health care and not believe in the hospital business model, which was driven by being unhealthy. That's why we felt that the focus should be more on primary care."
All of this makes sense and seems to be in theory the "intelligent" way to proceed. Numerous studies have identified clinical interventions that can be done to prevent hospitalizations. This includes tracking symptoms such as mental status changes and breathing difficulties, key measures such as blood pressure and blood sugar, and findings on the physical exam such as water retention and neurological changes and making lifestyle or medication changes in response. Garipalli emphasized that "the clinical interventions are all there. It's more about efficient and correct allocation to the right patients."
Of course, theory doesn't always match practice. Implementing this combined approach is not easy. Data always has its flaws. The human body and medical care are complex. You have to make sure that the machine learning algorithms don't generate the wrong insights. Good home-based care (and care in general) requires the right doctors, nurses, and other health care professionals. No amount of machine learning can compensate for a sub-par health professional. Ultimately, technologies can facilitate health care professionals and their decision making, not replace them. The day Dr. Ultron shows up at your door is the day to go hide and call the Avengers.
That's why validation of Clover Health's and others' machine learning approaches will be important. This entails demonstrating that such approaches are generating reasonably accurate insights and providing proper guidance. Ian Blumenfeld, Chief Data Scientist, and Brian Johnson, Senior Engineer at Clover Health, recently wrote that their "Hospital Admissions Data Model provides an 85 percent accuracy in identifying patients at risk for being admitted to the hospital in the next 28 days." Publishing such results in respectable peer-reviewed scientific journals can help engender more confidence in machine learning approaches.
Machine learning has grown in popularity in recent years, becoming the sexy cologne or perfume of health care. If you want to score points at a party, you can just say "I do machine learning to help do blankety-blank." There is certainly lots of potential, but just because you are using machine learning doesn't mean that you are doing something useful. You could even be making things worse. What's really needed are people, organizations, and approaches that can truly straddle the different aspects of health care, its financing, and technology to develop the right machine learning approaches and ensure that they are being used appropriately. Technology in itself does not have value. It's value depends on what real changes it can facilitate and what problems it can solve.
Using machine learning to tackle the preventable hospitalization problem can be a naturally intelligent thing to do. Again, if done appropriately. There are still far too many people getting hospitalized when such hospitalizations could have been avoided and people bouncing back-and-forth between their homes and the hospital. This is a failure of the current health care system and society. It is also a failure that people and traditional methods alone have not historically been able to solve. The system is too complex for traditional approaches, and much more learning of all kinds is needed, including perhaps machine learning.
