by Paul Bradley

Applications of predictive analytics in healthcare

Dec 22, 2015
AnalyticsHealthcare IndustryPredictive Analytics

Financial and clinical aspects of healthcare are inexorably intertwined under the broad umbrella of value-based care. This intertwining is by design, and is as evident at the macro and network-contracting level as it is at the microcosmic level of individual provider and patient payments.

health analytics ts
Credit: Thinkstock

Looking ahead at 2016 and beyond, two major applications of predictive analytics in healthcare stand out:

Financial performance management—the ability of healthcare organizations to navigate tumultuous market and regulatory shifts, and not be forced to scale costs exceeding their revenue.

Clinical performance management—the operational- and communications-associated processes that support and measurably impact outcomes and patient experience across the continuum of care. Population health management is a subset of this category.

Note that these two big-picture categories exist independent of care setting—meaning they are just as true for new-to-market retail healthcare clinics as for established hospital and health systems. Financial and clinical aspects of healthcare are also inexorably intertwined, for the foreseeable future, under the broad umbrella of value-based care. This intertwining is by design, and is as evident at the macro and network-contracting level as it is at the microcosmic level of individual provider and patient payments.

Let’s dig into these two major applications a bit more.

Financial performance management

If you know exactly what you’re looking for—and if the data elements that comprise these instances don’t change over time—traditional rules-based software can sometimes prove adequate.

If however, the data elements exist in massive data sets and are subject to change over time—which in turn drives instances or accounts into increasingly non-uniform iterations and combinations—static rules-based software quickly loses pace and lags further and further behind. Of greater concern is that manual review and analysis cannot identify and map the multi-faceted correlations that must be weighted in relation to each other before they become effectively actionable.

Massive…complex…iterative…comprised of elements that change over time—hey, that sounds like healthcare financial data. Consider merely a subset of a hospital’s to-be-filed claims—namely, those that are likely to be denied (for the purposes of this example, “likely” is understood to mean ≥90% probability of being denied).

There are all sorts of ways to slice this data—by dollar amount owed, by payer and insurance-plan type, by procedure code or frequency of appearance within the data set or even by the patient’s projected health system utilization. Yet the most impactful identification will always be one that takes into consideration metrics such as cost-to-rework/resubmit the claim, subsequent likelihood the claim will be paid (and projected timeline for payment to be received, which may vary significantly across payers), and the projected long-term modeled effect of that denial (and ones like it) on the provider’s operational and financial sustainability.

In the same way that these likely-to-be-denied claims and the patterns that cause them are difficult (in practical terms, impossible) for human beings to identify across millions of instances, so too are the missing charges, coding variances, and other meaningful anomalies that otherwise remain hidden in a hospital or health system’s financial data sets.

Predictive modeling technology, however, is capable of sifting through these massive data sets and uncovering the patterns and trends that correlate the target data elements to some outcome(s) of interest. Predictive modeling empowers organizations to ask fundamentally different questions than they can ask with software built of static, manually created rules.

For example, rather than asking, “How many claims exhibit these exact (and already known/defined) characteristics?”

We can instead ask, “what are the properties of a claim or of the information that goes into a claim that makes it more likely the claim will be denied—or not denied? What is the ultimate impact of that denial (and ones like it)?”

The end result is that organizations can be smarter about addressing the root-causes of those patterns upstream. One of the most exciting applications of this technology is contract modeling and management—healthcare contracts are becoming more and more complex, and as providers take on additional risk, they need technology that can effectively and rapidly model the impact of these new contracts based on myriad financial, clinical, and patient-population data elements.

Given that these contracts are the basis for how providers are reimbursed for the care they deliver, this is a natural segue to discussing clinical performance management.

Clinical performance management

Knowing when a patient is likely to shift to a less healthy, higher-risk category allows the provider to intervene to avert or delay the shift. The first step of modeling this likelihood is to uncover which previously unknown or untracked relevant patterns and commonalities in simultaneous, temporal, or aggregate combination create the greatest likelihood of a given specific outcome occurring (or not occurring) at the patient or patient-subpopulation levels.

Once these patterns are uncovered and subsequently validated, they become the basis for probabilistic modeling that can rapidly and continuously be refined both through machine learning and human instruction. These automated, proactive alerts are critical in maintaining and improving the health of a population of patients. Hospitals, health systems, and specialty as well as primary-care physicians are being asked to collect, record, store, and share exponentially more data than ever before—all in an effort to support overlapping initiatives of insurance reform, mandated health IT adoption, and state, federal, as well as payer-specific quality reporting programs.

Daunting as all of this sounds, we shouldn’t lose sight of an encouraging truth: providers already have data assets in the form of electronic health records and financial billing systems. Integrating these disparate sources together in patient-centered data sets provides the foundation for the application of predictive modeling to better understand patient populations. The resulting models are the core technology to compute and track the health and financial risk status of the patient population being served. This information is one of the cornerstones in attaining healthy outcomes while maintaining and reducing costs – keys for success in 2016.