The Healthcare Battleground + The False Narrative of Efficiency vs. Effectiveness

The brilliant team over at Axios recently dedicated an entire issue to the challenges facing our healthcare system. It is well worth the read and captures their facts forward approach to journalism. The underlying theme is that the healthcare system in the US is broken. Neither side, from single payer to full privatization has any supporting evidence to bolster their arguments.

The lack of consensus, even among unbiased practitioners argues for a radically different approach – one that allocates resources more effectively while retaining the core pillar of the doctor-patient relationship. To achieve order of magnitude improvements in resource allocation the industry is going to have to get much, much more effective at tapping its growing ocean of data. This precipitates the machine intelligence age for healthcare. At this point, it is not an if or a how, its a when and where.

The when starts now. Healthcare is the #1 issue in the minds of voters and Democrats will keep it there in the 2020 election. Medicare will be effectively tapped in 2026 meaning it will become THE #1 issue for the next President.

Any short term fix will focus on costs. This will precipitate a massive re-negotiation on reimbursement, squeezing profit margins across the board from the provider to the insurer and the drug maker. This will, in turn, kick off a massive lobbying effort as vested interests seek to retain their current stake.

That is the wrong where. The right where is allocating resources more efficiently. We know this as value based care. Aligning incentives with the effective and efficient deliver of care.

The ability to deliver against the goals of value based care is a function of two things – neither of which the industry does well today. The first is the ability to understand patient populations at a granular level. Second, is the ability to develop effective care plans for those fine grained patient populations. While seemingly simple, these are among the two hardest challenges in healthcare – having vexed the industry for the better part of a decade.

Over that decade, however, there were two important developments. Development one was the emergence of the EMR and with it an extraordinary trove of data on patient outcomes, costs and treatment approaches. Development two is the emergence of new analytical techniques – namely machine intelligence. Using these techniques, healthcare can fulfill the promise of value based care by first understanding population risk and then creating multi-factorial care process models for those populations.

More importantly, this is already occurring in practice by innovative players on both the payer and provider side (or both in the case of an IDS). Further, it is happening at giants like Intermountain and at community hospitals (stay tuned to learn who).

These are data problems first and foremost. It doesn’t diminish the accompanying process changes, but the industry has what it needs to become as far, far more efficient – almost overnight.

Let’s start with the challenge of population risk stratification. For too long, payers and providers have viewed their populations through the lens of monolithic disease states. One reason is that traditional analytical techniques are not able to present a clear, justifiable picture of the multifactorial nature of a system’s highest utilizers. Because we think of healthcare in terms of the most chronic presenting disease state, we tend to treat and view healthcare through that same lens – missing the fact that these high utilizers are that way precisely because they have multiple co-morbidities.  

New techniques come at the problem differently. The emergence of unsupervised learning allows payers and providers to find the patterns and relationships that exist in the data without having to have an outcome variable in mind. By not asking a single question (thereby eliminating others), one opens up to the possibility of any number of questions by finding the underlying structure of the data. What emerges from an unsupervised learning approach is a far more fine grained understanding of the patient groups and collections of conditions.

While grossly oversimplified the following example outlines how an already innovative IDS applied these techniques to their population risk challenges. This IDS was working with standard monolithic definitions – let’s call them Cancer, Diabetes, Chronic Pain, etc.. They knew, and had identified, that there were multiple subgroups of these disease states. Cancer, for example had multiple subtypes that required different care: breast cancer and prostate cancer.

What they also knew, but struggled to find in their data, was that prostate cancer often had accompanying co-morbidities and that they were not accounting for that in their care plans as well as they wanted to, or in their financial projections.

By using analytical techniques like Topological Data Analysis, which combines both unsupervised learning and supervised prediction, the IDN was able to identify the key populations of prostate cancer + other co-morbidities.  These smaller patient groups provide for significantly better cost prediction – more than 50% better in than traditional techniques.

The prediction of costs, while important, is not as important of delivering better care to those patients. The multifactorial nature of these patient groups represent a particular challenges for care process model development. Again, because medicine has tended to think monolithically with the most urgent presenting condition, care process models are also monolithic in nature and tend to focus on acute care (total knee replacement, colorectal, CABG). Here again, using unsupervised learning to identify the factors that contribute to better outcomes for similar patient groups – we now have the ability to develop multifactorial, human understandable consensus care paths. This leads, by definition, to more complex and generally longer care process models, but also to better patient results.

This is important because the technology to view the patient in this multi-dimensional space is not enough to deliver against the triple aim. Rather, payers and providers need to embark on a transformative journey to alter their processes, their training and their delivery networks to deliver against this new view of the patient.

Google’s ambition’s in this area are predicated in part on this unsupervised approach – because they have the capabilities as well realization that iteratively testing hypothesis won’t get us where we need to go – at least not anytime soon. And time is running out.

The challenges precipitated by the clock will spur real change and this is good for patients, providers and payers as they finally unlock the massive data investments and finally delivering against the promise of value based care.