In January 2015, the U.S. Department of Health and Human Services (HHS) boldly announced a plan to tie 30 percent of traditional fee-for-service, Medicare payments to quality or value through alternative payment models such as accountable care organizations (ACOs) and bundled payments by 2016, and tying 50 percent of payments to these models by the end of 2018. HHS also set a goal of tying 85 percent of all traditional Medicare payments to quality or value by 2016 and 90 percent by 2018 through initiatives such as the Hospital Value Based Purchasing and the Hospital Readmissions Reduction Programs.
But, earlier this year, a survey from Salt Lake City, Utah-based analytics vendor Health Catalyst revealed findings that many expected: most industry stakeholders seem to think the government was quite ambitious with these projected numbers. The survey, at the time of its publication, found that just 3 percent of health systems have already met the target set by HHS. Only 23 percent expect to meet it by 2019, just a year after the feds had hoped that half of all Medicare reimbursements would be value-based. What’s more, the majority of health systems—a full 62 percent—had either zero or less than 10 percent of their care tied to the type of risk-based contracts identified by HHS as “value-based,” including Medicare ACOs and bundled payments, the survey revealed.
The healthcare executives surveyed did say that they intend to steadily increase value-based care and at-risk contracts, and they said the most important organizational element needed for success with risk-based contracting is analytics. This is where Leonard D’Avolio, Ph.D., an assistant professor in the Brigham and Women’s division of general internal medicine and primary care, says change is needed. Dr. D’Avolio is also the CEO and co-founder of Cyft, a company based on years of his research optimizing machine learning and natural language processing to improve healthcare. He previously led informatics for the Department of Veterans Affairs’ (VA) precision medicine initiative (the Million Veteran Program) and the first clinical trial embedded within an electronic medical record (EMR) system.
D’Avolio fully understands that the success of value-based care is dependent on healthcare stakeholders understanding and predicting what will happen based on the information they have. Thus, he recommends a different approach to analytics from what has traditionally been practiced in healthcare. He says, “As value-based care organizations are now discovering, these multi-million dollar investments in traditional analytics are useful for understanding what happened—how many beds were filled, drugs prescribed, surgeries performed. However, they are incapable of answering the fundamental questions of value-based care: what should happen, to whom, when, and how, in order to prevent future events.”
As such, he says that most clinically relevant information is ignored by traditional analytics. To this end, as part of Healthcare Informatics’ Special Report on data analytics in this issue, D’Avolio recently spoke to Managing Editor Rajiv Leventhal about what needs to change in approaches to leveraging analytics in healthcare’s value-based future. Below are excerpts of that discussion.
Can you tell me a little about your company, as it relates to the future of healthcare, and healthcare analytics?
Our company is focused on making technologies—such as machine learning and natural language processing—available to data analysts so they can harness the power of predictions in ways they haven’t been able to. We try to find organizations where the chief financial officer and the chief medical officer have the same incentive, meaning the organization is at financial risk for delivering high quality care. Frankly, relatively little of care provided at hospitals is at true financial risk today, though that number is increasing. Most companies are incentivized to still invest in technology to help them see folks more quickly. We are happy to see that changing though.
Sure you can talk about readmissions, but when you are at full financial risk, what you really care about is preventable utilization. Our customers will sometimes start the conversation asking about readmissions, but we ask them, what interventions do you have at your disposal? They might say that they hired a nurse to focus on COPD [chronic obstructive pulmonary disease]. So we say to them, what if we build a model to identify exactly who in your COPD population will end up in the ER in the near future? It’s a different approach from today’s risk scores, which is limited to claims data and is too one-sized-fits-all, with a focus on only a few problems. These approaches treat the geriatric patient with heart disease the same way as the high-risk pregnant patient. So we are trying to move away from one-sized-fits-all approaches.
Leonard D’Avolio, Ph.D.
How do you view the overall landscape in terms of analytics being leveraged by payers and providers as they move into risk-based contracting and reimbursing for value rather than volume?
The writing is on the wall; unless there is going to be a major political shift that comes with it a gutting of CMS [Centers for Medicare & Medicaid Services] policy, we are moving towards value-based care in various forms. CMS opened Pandora’s Box leading with ACOs, alternative payment models, and bundled payments, and the commercial plans have been waiting for this forever. When CMS fired that first shot with ACOs, many of the commercial plans turned to ACQs, or alternative quality contracts.
If you are having to do ACO models as part of your CMS reimbursement anyway, why not make it even more attractive and easier by giving more flexibility and creating alternative contracts so you can go at risk with us also? Most of the fee-for-service world is now keeping a close eye on MACRA [the Medicare Access and CHIP Reauthorization Act of 2015]. With MACRA, CMS and others drew a line in the sand saying that we will be more than 50 percent value-based by 2018.
One of the major challenges in value-based care is that we are in that quantum state in healthcare where there isn’t a value-based care policy; even the ACO program has many different reimbursement and quality measurement policies. There are a number of things with alternative payment models that need to be measured and reported on, too. It’s not getting talked about much, but healthcare is not transitioning to one new way of paying for care. In fact, depending on how you measure it, there are between five and 12 versions of this, many being implemented at the same time by provider organizations.
This leads to challenges around analytics for IT departments. With each new flavor of reimbursement usually comes a new layer of process measures that needs to be reported against, which usually means new bolt-ons to the EMR, which was never designed to improve the quality of care to begin with. So you are taking an EMR, which was designed 30 years ago for financial reimbursement, to communicate narratively between clinicians, and to ensure legal protection for the [provider], and you now bolt on dozens of new process measures. So you’re doing this quantum value-based care transition, and that creates challenges.
So what are the best analytics tools out there today?
In order to use analytics successfully, you want to take all information and turn it into actionable insight based on the organizations’ own highest priorities. Now, there is branching going on with analytics, driven by financial incentives. There are two branches that analytics are forced to operate within, and one of them is mandated reports based on each of your payer contracts. These are just reports, and they are mostly designed around the things that both sides agree on in advance and can probably be done based on using the EMRs we already have. The problem with mandating reporting is that we’re doing the opposite of what led to the digital transformation as experienced in other industries.
When other industries became digital, they had agreed upon outcomes, but then the competitive advantage came when they used all of their data to discover the best way to get to those outcomes. Amazon and Netflix, for example, did this by learning everything about the consumers they were serving. That’s the competitive advantage—taking all of the data and then becoming very personalized towards the recommendation and an agreed upon outcome. Healthcare has done the opposite in this branch of analytics—which is take all of the data we have, only look at a few points in time, create the standard patient and standard workflow, and somehow people think that will lead to the desired outcome.
The second branch of analytics is about organizations discovering the most efficient ways to do things. Because now, for the first time, you have to be able to make sense of all of the data, and you have to prioritize it for the care delivery folks. You can’t tell them that readmissions matter most to you. Instead, you have to say, “Here is the outcome—improve care—now you have to use your data to figure out the best pathways to get there.” In effect, you are becoming more like every other industry, in which digitization can reach its full potential. So I think if this happens, 95 percent of what passes as analytics today will either become obsolete or change dramatically.
Drilling down, regarding CMS’ readmissions reduction program, and the government’s mandatory bundled payment programs, how can I.T. leaders better use data analytics as they participate in these processes?
If you are going to work in the bundled payment world, you need to be able to anticipate and not react after the fact. So you have to become much narrower in your predictions. It’s not just about readmissions, but for example, which of my patients is most likely to end up with a non-routine discharge so I can begin to prepare that patient in the most cost effective pathway possible? That’s a specific example of something we are doing now, and we are finding that you need to be able to consume far more than just claims data in order to make those kinds of predictions. Any tool that uses claims data alone has one arm tied behind its back. The claims data is dated, and also, ICD-9 codes can be up to 70 percent inaccurate depending on the disease. This will only get worse as we move to the 65,000 disease codes of ICD-10.
Another thing we are involved in now that you wouldn’t think about in a traditional value-based sense is patient satisfaction. We’re working with a managed care organization around which members of their population are most likely to disenroll after being in the program for one year. From a CMO and CFO incentive perspective, if you are going to invest in keeping folks healthy for a year, and take on acquisition and health maintenance costs, then knowing who will leave after a year is a big deal. So we are taking on 30 different file types and helping this managed care organization predict who is likely to leave in a year. It’s not curing cancer, but it’s critical for organizations to survive. People will have to get more granular using all their data rather than one-sized-fits-all risk scores for readmissions.
What advice can you give to CIOs, CMIOs and CMOs as they continue to prepare for this new world?
First, analytics is not a tool; it’s a process. Clinicians understand where to focus, but you need to come up with the processes, tools, and support staff that will help and empower them to identify the highest priorities. Also, measure them on where it’s working and not working with ongoing feedback loops. It’s a problem to think of analytics as a product that you buy that will lead to behavior change, workflow change, and process change.
Second, be able to distinguish between what counts as analytics in the fee-for-service world with what will be required of analytics in a value-based world. You need to move beyond claims data and really use all of your data. It’s about understanding not just what happened, but what is most likely to happen and what you should be doing about it. This is very different than the traditional approach of what is considered analytics in healthcare.
Third, no CIO should settle for a vendor’s insistence about what’s good enough when it comes to predictions. If you are building models based on other peoples’ data and other peoples’ priorities and populations, you cannot presume that can be brought into your shop and will perform at the same level. CIOs need to own the evaluation with their own data on their own problems.