To better understand the fundamentals of predictive analytics — and why it has the potential to transform healthcare — it can be helpful to use Netflix as an illustrative example.
Let’s say, for example, you’re sitting around the house one rainy October Saturday and decide to view a few movies using Netflix’s streaming service. First you watched Field of Dreams. Then you decided, hey, this rain isn’t letting up any time soon, and that dog doesn’t want to go out in it any more than I do. So, after a brief backyard sojourn during which you and the dog confirmed that 38 degrees and rainy is in fact unpleasant — you reconvened your Netflixing and ended up watching Bull Durham, as well. Further, let us also assume that you enjoyed both movies and watched them all the way through.
As the credits rolled on Bull Durham, the critical question for Netflix was the same it always is: What would you enjoy watching next — specifically, what should Netflix recommend? Based upon the day’s viewing you may have a soft spot for baseball movies. Though it could just as easily be the case that you’re Kevin Costner’s biggest fan and the fact that you queued up two of his baseball movies was pure coincidence.
Given the uncertainty orbiting these pieces of information, maybe the best prediction would be Dances with Wolves, starring Kevin Costner. Or maybe the right pick would be Moneyball, the story of how the Oakland A’s leveraged data-driven, evidence-based sabermetrics to remain competitive against much more highly capitalized MLB teams. But what if neither Kevin Costner nor baseball is the most important correlation—what if the best predictor of whether you’ll like a film is simply whether it’s a sports movie from the late 80s?
As with all forms of predictive analytics, the question of what to recommend multiplies in complexity as overlapping variables (often in the form of unstructured data) are added and subsequently considered within algorithmic equations that power, in this case, Netflix’s recommendation engine. Further complicating the matter, it’s likely that you’re not the only person to have watched both of these films in close proximity and there are likely to be numerous “motivations” for such viewings across the population. It becomes apparent rather quickly the inherent challenge of something that seems, on the surface, as straightforward as a recommendation engine.
In healthcare we face these same kinds of challenges, just in a different form. The questions we ask are which gaps in care create the greatest risk for the patient, or which specific combinations of gaps in care correlate with readmissions—so that clinical outreach coordinators and other staff can prioritize whom to contact right away. We ask which types of claims are most likely to be under-coded or missing charges—so that organizations can make best use of finite resources like staff time and ensure the greatest positive impact on overall financial performance.
Questions in this vein also form the basis of accurate risk stratification and effective disease management. They enable healthcare organizations to distinguish between seemingly similar patients within disease registries—patients who might at first or even second glance appear to be equally at-risk or in need of care intervention, but who are on distinctly different paths likely to lead to dramatically different clinical outcomes.
More of the Right Things Brought to Your Attention
In Netflix’s case, predictive analytics works when you discover movies or TV shows you enjoy.
In healthcare, it works when you identify patients who share specific risk factors or profiles even though they may not share obvious or previously flagged commonalities; it works when you zero in on the subset of claims that are missing codes or other critical information that you don’t even realize is missing—and that is negatively impacting your reimbursement.
Let’s say you have two patients who are both at risk for developing complications related to diabetes and who, at first glance, haven’t been in for their blood-work and other screenings in over six months. By bringing in additional data points and assigning those data points the correct significance relative to the population, you can see that one patient likely hasn’t been in for their appointment because their health has improved thanks to rigorously following other aspects of a disease management program—while the second patient is likely experiencing a decline in health as evidenced by new claims for durable medical equipment fulfilled through another provider.
Predictive analytics also reduces the risk that healthcare organizations will waste resources manually triaging, prioritizing, or qualifying the information that comes into their systems or work queues—processes that in most cases can be partly if not fully automated. This in turn reduces the risk of wasting resources trying to close gaps in care that aren’t as critical to improving patient outcomes—or, to use an example from the RCM world, reworking claims that aren’t that impactful to your organization or your bottom line.
Getting smarter as you go: that’s challenging. Who’d have thunk?
To ensure the baseline accuracy of predictive analytics continually increases, the algorithms and layers of artificial intelligence at work must incorporate new information, determine its degree of relevancy, and then—if it meets at least one of multiple independent thresholds for correlation or relevancy—accurately weight the new data in relation to historical norms.
What happens when there are too many variables to make sense of—or when they aren’t properly weighted?
Well, in the world of streaming video service, your “Recommended for You” queue might start to include films that are tagged as inspiring, family-friendly stories of overcoming adversity. These would no doubt be rich musings on both the pastime and the passion of baseball, sonatas of the American experience itself, timeless works that explore what a father carries with him onto the baseball diamond—and what he must set aside if he is to focus on the game.
You know, like Air Bud: Seventh Inning Fetch.
Shifting our focus back to healthcare: Inadequate predictive analytics could result in under-identifying or misidentifying claims that should be flagged for a set of human eyes to review—all because the algorithm ascribes too much importance to the lack of (or presence of) a specific data point.
In the same vein: If the algorithm can’t effectively distinguish between relevant instances and instances that share non-relevant commonalities, the resulting dataset might turn out to be too large or too messy to be of any real use. Rather than sharpening your focus, you might end up manually working through what amounts to a random subset of data—no real improvement over spot-checking.
All these scenarios are different symptoms of the same problem: The algorithm assigning too much or too little importance to the wrong data points.
The bigger picture — because healthcare data will only get bigger
Predictive analytics doesn’t just make mining big data less prone to human error—it also supports making the best use of human time and effort, because it can reliably flag the things that a set of human eyes should review. Whether it’s powering Netflix’s user experience or enabling more focused and results-driven financial and clinical performance management, what it’s really predicting is the same thing:
What you want to see — and need to see.
On Netflix that might come in the form of sports movies or one of the innumerable Kevin Costner classics.
In the increasingly interconnected worlds of population health management and revenue cycle optimization, it comes in the form of reporting that identifies the instances that matter—the ones that have the greatest impact on your clinical outcomes, operational efficiency, and overall financial performance.