Using data to identify trends and patterns can help drive better outcomes.
At a Glance
Predictive analytics can be used to rapidly spot hard-to-identify opportunities to better manage care-a key tool in accountable care. When considering analytics models, healthcare providers should:
- Make value-based care a priority and act on information from analytics models
- Create a road map that includes achievable steps, rather than major endeavors
- Set long-term expectations and recognize that the effectiveness of an analytics program takes time, unlike revenue cycle initiatives that may show a quick return
Accountable care organizations (ACOs) will come in many configurations over the next few years, but one underlying strategy will prove indispensable in any configuration: making use of actionable data to drive and sustain improvement.
In the new value-over-volume healthcare model, as the Centers for Medicare & Medicaid Services (CMS) prepares to reduce payments to providers with higher-than-expected 30-day readmissions for certain conditions, such as pneumonia, providers will need to target care that is not cost-effective. Simply put, CMS will no longer pay for wasteful care. But how do hospitals determine whether care is wasteful?
The answer can be found among the pieces of patient, clinical, and financial data spread across various hospital information systems. ACOs must collect, integrate, and analyze these data to fulfill their fundamental purpose of reducing costs by providing better care for specific patient populations, such as patients with type 2 diabetes.
Reaching this endpoint, however, requires more than just analyzing integrated data. A 2011 report by the American Hospital Association, Hospitals and Care Systems of the Future, says hospitals will not only need the right kind of technology, but also must be able to extract useful metrics that can be used to meet the requirements of the new value-based model. According to the report, "The ability of an organization to leverage the technology to perform sophisticated data mining and analysis in real time for continuous care improvement is critical for long-term organizational sustainability."
The key to population health management, therefore, is optimizing data by extracting patterns and correlations using "sophisticated data mining," or predictive analytics, to identify care practices that result in better outcomes, such as reduced readmissions.
For hospitals, distinguishing cost-effective from ineffective care not only will reduce healthcare costs overall, but also ultimately will protect revenue.
The good news: Hospitals already store and collect an enormous amount of patient, clinical, and financial data. Among health information systems, electronic health records, cost accounting systems, clinical systems, and chargemasters, terabytes of valuable data are generated and collected for analysis.
The bad news: Making use of such information is challenging. Such systems often support operational objectives, such as generating and supporting claims, and act as communication devices, but not as analytic or business intelligence tools. HFMA's 2011 report Value in Health Care notes that although such business intelligence may be the most important of four capabilities that organizations require to prepare for value-based payment, few organizations have the expertise or resources to develop such systems. Many hospitals simply lack the infrastructure necessary to build a data warehouse that acts as a sufficient platform for predictive analytics.
In its basic form, a data warehouse is a repository of data from disparate information systems. Despite being in one place, the data are raw and not adequately prepared for use in meaningful, reliable metrics.
In an effective data warehouse, the data are prepared to allow for easy access and manipulation to build analytic models that support accurate and reliable conclusions.
To build such a data warehouse, the data must undergo a three-part process.
Staging. Data that are most useful in drawing patterns and correlations are pulled from each source and prepared to be imported into the data warehouse. For example, important information for data mining and predictive modeling is provided via the procedure codes, diagnosis codes, and financial and clinical data associated with an account. The staging phase of data warehouse loading includes identifying sources for diagnosis codes, procedure codes, financial data (e.g., charges and payments), and clinical data associated with accounts. Preparation of these data consists of validating that codes are properly ordered and the financial data is in the correct format.
Integration. Pieces of related data from disparate sources are aligned and standardized. All data related to a specific patient, for example, are linked to that patient. Attributes, such as payer names, are made consistent to ensure the integrity of the data.
Access. The data are structured to allow for easy and timely access from the data warehouse to support scalable analysis, modeling, and prediction. For example, consider a report that provides management with a view of total account charges and payments by procedure codes, diagnosis codes, and financial and clinical data elements associated on the account. The data warehouse can support easy and efficient access to these data by joining the following information: account charges and payments aggregated for each account; procedure codes, diagnosis codes, and clinical data elements associated with the account; and the date in which service was given. This report requires joining financial data (charges and payments), procedure codes, diagnosis codes, and clinical data elements associated with the account and a common date format across these sources. This structuring will allow a report to be generated and used across the organization.
Neglecting any of these steps when building a data warehouse could result in incorrect analysis and spurious correlations-making effective decision-making impossible. Even worse, an inadequate data warehouse may produce erroneous predictions, resulting in faulty analysis or ineffective care.
Digging Deeper into the Data
The data warehouse provides the primary source for identifying patients with specific conditions that are particularly important targets for predictive analytics, such as type 2 diabetes and pneumonia.
Predictive analytics, then, dig deeper into the data to identify trends and patterns that are not readily apparent in the data at face value. For example, predictive analytics can detect patterns among patients with type 2 diabetes by sorting through thousands of attributes (e.g., number and type of patient visits, types of insurance, charges, admit locations, diagnoses, and procedures). As a result, similar subpopulations of patients with type 2 diabetes are discovered and correlations and trends indicative of various outcomes become apparent.
These subpopulations (or clusters) of patients are then used to identify, for example, patients whose conditions are being effectively managed, patients most at risk for developing a worsening condition, patients with the greatest number of inpatient and ED visits, and patients who are likely to be readmitted, among other outcomes.
The value of the data is increased by deriving additional data attributes, allowing outcomes to be identified with increasing accuracy by means of improved predictive analytics algorithms. For example, hospital systems will track a patient's discharge and readmission dates, but not the time between. Calculating the number of days between discharge and readmission provides an important additional piece of information that is used by the algorithms to increase the accuracy of predictions for specific patients.
By highlighting correlations and patterns among data, predictive analytics connect financial and clinical outcomes to help identify the care processes and services that produce the most effective results. The tool ultimately helps financial and clinical leaders distinguish between effective and ineffective spending.
Predictive analytics can be used to drive performance in two areas where ACOs will be held accountable: reduced readmissions and preventive care.
Readmissions. Predictive analytics could predict the likelihood of a patient being readmitted for the same condition, which is especially important because CMS penalties for 30-day avoidable readmissions begin in 2013. Predictive analytics find correlations among patient account, charge, and clinical data elements to determine the key attributes that will most likely result in a readmission versus attributes that will not likely lead to a readmission.
For example, data elements for patients with pneumonia collected over a 12-month period at one health system included patient demographics, procedure and diagnosis codes, CPT/HCPCS charges, and revenue and DRG codes. The data elements were augmented with derived attributes, including time between discharge and readmission, length-of-stay groupings, and age groupings. Predictive analytics algorithms sifted through thousands of data attributes, analyzing data on patients with pneumonia who were readmitted and those who were not. The results showed that 26.2 percent of patients who were readmitted had a code labeled "patient refused therapy statistical charge," while 21.6 percent of those who were not readmitted had an endotrachial tube procedure.
To actually reduce readmissions, the models constructed by the predictive analytics algorithms use each attribute to attach a risk score to each patient before discharge. This score helps in determining whether additional treatment and a longer stay are required to reduce the likelihood of a readmission. For example, predictive analytics produced models for the previously cited health system that, for a subset of the Medicare population, could identify pneumonia readmissions with 75 percent accuracy, representing a 217 percent improvement over the health system's accuracy rate prior to using predictive analytics.
Preventive care. The fundamental steps in providing preventive care are identifying patients with existing conditions that may lead to more severe conditions and taking action to avert such outcomes. For example, the case study health system used predictive analytics to identify similar subpopulations of patients with type 2 diabetes. The algorithms identified 10 such subpopulations by sifting through thousands of patient attributes collected over multiple visits in a 12-month period. The resulting cluster model defined patients, for example, with low cost and low number of visits, medium cost and frequent outpatient visits, and highest cost per visit.
One of the clusters identified costly infrequent visitors-younger patients (average age 54) with a low number of total visits (on average, 1.73 visits in 12 months), but with an average cost per visit of $23,760, approximately 3.6 times the average cost for the entire population. These patients had a low number of total inpatient, outpatient, and ED visits, but these visits required a higher level of care and were more expensive.
In this way, predictive analytics allows high-cost populations of patients with type 2 diabetes to be compared with healthier, lower-cost subpopulations. Providing these patients with more frequent lower-level care, such as regular education on the importance of a proper nutrition and getting regular checkups, could prevent the condition from worsening, thereby avoiding more expensive care.
The health system also used predictive analytics to group patients diagnosed with morbid obesity. Statistical tests resulted in seven distinct subpopulations. Analysis of average total visits, total inpatient, outpatient, and ED visits, total charges, charges per visit, and average patient age across clusters identified patient subpopulations with the fewest visits and lowest cost and patients with moderate visits and high cost.
Another cluster identified patients with few visits (just 1.55 over a 12-month period, as compared with 3.41 for the total population), but with a high cost per visit (averaging $15,460, or more than twice the population's average charge of $7,685). As with the analysis of patients with type 2 diabetes, this subpopulation had a low number of total inpatient, outpatient, and ED visits, but these visits required a higher level of care and were more expensive.
In each of these ways, predictive analytics can be used to provide actionable information to users rather than forcing users to hypothesize and pull information. Cost data are linked with clinical data to pinpoint which patients require more attention to prevent their conditions from worsening over the long term.
Rewards and Risks
Healthcare organizations can use predictive analytics to identify at-risk patients in attempts to prevent readmissions and to find high-cost patients who may benefit from preventive care. They can also use predictive analytics to determine which procedures and care methods are likely to produce the most effective outcomes.
By providing meaningful data, predictive analytics give hospital administrators and clinicians information to make more informed value-based decisions to effectively manage patient populations. As the technology identifies care practices and patterns that result in better outcomes, quality of care improves, too.
For example, predictive analytics can help hospitals develop care protocols for chronic disease management, which can reduce readmission rates and mitigate the risk of Medicare reimbursement cuts for ineffective care.
But implementing ACO analytics is not a panacea. First, clinicians must interpret the data and decide whether to take action. There is no guarantee that their decisions will be infallible.
Second, it's not enough for clinicians to act. The patient must act, too. Physicians and healthcare staff can do only so much to influence patient behaviors. Predictive analytics can uncover services that are least effective and behavior that can lead to improved health, but they can't force patients to exercise regularly or follow other care protocols.
As providers become more accountable for care, caregivers should persuade patients to assume greater responsibility for their own health to improve care and reduce costs. Hospitals need to mine their data to ascertain which care is most cost-effective. Their bottom lines depend upon it now more than ever.
Key Considerations for the Analytics Model
Healthcare organizations that are contemplating implementation of a predictive analytics model should keep in mind four key considerations related to such an effort.
First, the organization should make value-based care a priority, from clinical care staff to finance executives. Information from analytics models should be acted upon or used as the basis for decision making-from physicians and nurses to care coordinators and patient safety clinicians.
Second, the healthcare organization needs to invest in the appropriate hardware, software, and dedicated IT staff to build a data warehouse or work with a predictive analytics/data mining solution provider.
Third, when employing ACO analytics, it is especially important for the hospital to have a road map that consists of achievable steps, rather than major endeavors. For example, rather than integrating predictive analytics with an electronic health record system at the start, simple alerts can be sent to flag areas requiring focus, including patients who are likely to be readmitted or patients who may benefit from preventive care.
Fourth, although the hospital needs to measure analytics implementation to gauge ROI, the hospital also should understand that the effectiveness of an analytics program takes time, unlike revenue cycle initiatives that may show a quick return. For example, interventions designed to prevent a patient with morbid obesity from developing type 2 diabetes will not provide immediate results. Obtaining results requires tracking patients over time to measure whether that proactive care results in improved health and fewer inpatient and emergency department visits.
Paul Bradley, PhD, is chief technology officer, MethodCare, Inc., Chicago, and a member of HFMA's First Illinois Chapter (Paul@methodcare.com).
Publication Date: Monday, April 02, 2012