Fundamentals of Effective Cognitive Predictive Analytics

May 2, 2017 11:44 am

Cognitive predictive analytics can prove a useful tool as health care enters an era where high quality is rewarded over high volume.

Thanks to alternative payment models that offer providers incentives for delivering high-quality care, the idea and measurement of value are changing in health care. The goal of the ongoing shift away from fee-for-service models is to promote improved care for individuals and populations at a reduced cost. Most within the industry agree that giving providers incentives to achieve better outcomes is a good thing. How best do so, however, is a major area of debate.

When viewed from a cruising altitude, the emerging value-focused geography looks like a patchwork of programs, mandates, and incentives that link provider performance with payment. This movement started in 2008 with the Medicare Improvements for Patient and Providers Act (MIPPA). The programs enabled by MIPPA and subsequent legislation, including the Affordable Care Act (ACA), are focused on some variety of payment model in which penalties and/or incentives are issued based on target performance measures. These programs, all administered by the Centers for Medicare & Medicaid Services (CMS), embody the multiple levers across standards, networks, reporting, certifications, and other innovations that underpin CMS’s quality strategy.

The common thread to most CMS quality programs is prevention. For example, the Hospital Readmissions Reduction (HRR) program focuses on preventing readmissions of patients as inpatients, and the Hospital-Acquired Conditions (HAC) program aims to reduce the incidence in hospitals of events such as pressure ulcers and surgical site infections.

The problem with mandates as a stand-alone approach to driving prevention is that they provide incentives without delivering the tools providers require to prevent the targeted events and improve health outcomes. The solution for many providers has been to turn to technology solutions, most notably predictive analytics, to help address the mandates.

Predictive analytics is a concept that has been adopted to describe a category of solutions. The specific predictive analytics solutions referenced here, cognitive machines, use deep machine learning (i.e., artificial intelligence and cognitive science) to account for thousands of variables, analyze vast numbers of patient dimensions (in the billions), and deliver actionable information for setting priorities and making recommendations. Outside of health care, cognitive machines are widely leveraged because they produce valuable results and highly specific insights almost immediately. For instance, cognitive machines enable internet service providers to know which ads to target to specific consumers based on search habits, location, and even the type of browser used (e.g., targeting 40- to 50-year-old men in the Atlanta suburbs with lawn mower ads in April).

Cognitive machine learning technologies have been around for years and leveraged by other industries from consumer goods to financial institutions with great success. It is time for health care to begin adopting these capabilities to drive value and improve patient lives.

Craftsmanship Counts: Qualities of an Effective Solution

Not all predictive analytic solutions deliver the capabilities inherent to cognitive machines. Some require the establishment of an expensive data warehouse, others may only work for very specific populations, and still others are not truly predictive. A provider seeking a solution that accurately and effectively targets at-risk patients, performs effectively in a clinical environment, and drives speed to value should ensure that the solution exhibits the following five key attributes.

Effectiveness. The most effective solutions identify at-risk patients more precisely than statistical models or traditional prediction tools such as the LACE index, which bases its predictions on four, static factors: length of stay (LOS), acuity of admission, comorbidities, and emergency department (ED) visits. The cognitive machine should get “smarter” about a healthcare organization’s patient population over time. Every new piece of data should add to the machine’s understanding and help localize outputs so that the recommended actions are more effective and patient specific.

Applicability. Sometimes it is hard to determine which problem to tackle first. As a general rule, it is best to target scenarios that are high risk and high dollar, where interventions are easily applied and require low investment. Many of the diseases and conditions targeted within CMS’s quality strategy qualify.

Based on 2013 data, the average cost of a Medicare readmission is $13,100, and the cost of treating a hospital-acquired stage IV pressure ulcer can approach $130,000. a The interventions to prevent these types of events are relatively easy to apply, minimally invasive, and cost effective. An effective cognitive machine should include vectors built around such events to enable an organization to drive effective clinical action (e.g., schedule a follow-up appointment with the primary care physician) with minimal development time.

Optimization. A defining characteristic of cognitive machines is that they can go beyond predicting the likely incidence of illness within a population to provide operational insights that will enable organizations to better allocate their resources for treating the illness. In short, the solution should quickly and easily deliver outputs that identify areas of risk, predict when those risks will occur, and provide insights on the best interventions that will engender engagement.

Efficiency. Machine learning technologies are mature enough that they can use very little and very “dirty” data (i.e., data that are disparate, incomplete, and possibly inaccurate) to deliver accurate predictions. Missing data complicate a model’s ability to differentiate between a negative value and incomplete information. For example, a person without a documented history of hypertension may or may not have the condition, but if he or she did have it, the clinician didn’t include it in the patient record. Using deep machine learning capabilities that apply advanced mathematical principles, an advanced, machine-learning-driven cognitive solution can compensate for such data shortcomings to deliver the most effective predictive outputs.

Usability. Usability relates to how easily staff can implement the insights and recommendations produced by the machine. In other words, a cognitive machine’s usefulness depends on how easily the patient insights and recommendations it produces can be incorporated into the organization’s workflow. The real differentiator with any kind of cognitive science, therefore, is that the extent to which it can promote greater effectiveness (i.e., better allocation of resources, ability to more precisely identify individuals, and recommended actions that have the greatest likelihood of mitigating risk). It then is simply up to the clinician just needs to use the outputs. Consider, for example, the difference between an X-ray and an MRI. The information from the MRI is so much clearer, more precise, and actionable that it reduces the cognitive load required to interpret outputs and take effective action.

Most of the power within a cognitive machine sits well out of sight of the end user. The engine that drives the associations, prioritizations, and recommendations is invisible, in that the user does not see the data and calculations it makes to produce its outputs. Like advanced search engine technology, outputs are embodied as smart, prioritized lists of recommended actions. These lists are then inserted directly into the established clinical workflow to ensure adoption. The nature of the machine facilitates perspicuous, patient-specific recommendations that are easily consumed by varied user interfaces.

For example, one southeastern hospital uses cognitive machine technology to identify patients who are at risk of pressure ulcers. Risk and intervention information is inserted directly into the facility’s electronic health record system and rendered at the point of care. The result has been a 33 percent drop in pressure ulcers within the first few months of workflow integration.

Defining the Target

By their nature, cognitive machines are designed to help providers stop patient deterioration and/or illness by enabling early intervention. The challenge is that the process of applying an intervention obscures the value of the prediction.

In other words, by applying an intervention, a healthcare organization loses the clear-cut ability to measure—at the individual level—if its actions prevented the illness. This phenomenon can be called the predictive analytics paradox.

For example, a high-risk patient receives a flu shot with the goal of preventing the flu. If that patient does not get the flu, there’s no way to know whether the vaccine was responsible for the outcome or the patient was not going to get sick in the first place.

At the individual patient level, we can’t clearly determine if the intervention translated into value. We continue to vaccinate patients because we know that at the aggregate level, vaccinating against the flu lowers the occurrence of the virus. It is this same logic that we apply to predictive analytic solutions to assess its effectiveness. The target in this scenario isn’t one point that encompasses the solution’s value; it is the aggregate of points and the comparison of performance across time.

Hitting the Bulls-Eye

Implementing a cognitive machine is relatively easy. And if done correctly, the machine can start to deliver value at the earliest stages of end-user adoption. The concept behind a cognitive machine is that it is “plug-and-play.” A provider feeds data into the machine, the machine analyzes these data across billions of factors, and the machine delivers outputs that can be inserted into the existing workflow and preferred user interface (e.g., data visualization tools, electronic health records, patient management software). Users in the earliest stages of adoption and engagement receive information that helps them focus resources and clinical action so that savings are quickly realized (e.g., the aforementioned average of $13,100 tied to readmissions, reduced length of stay), resources are optimized, and quality is improved.

As an example, using cognitive machine technology, a large integrated delivery network has saved close to $2 million in material costs related to readmissions. The machine also serves as the driver for all care coordination and has helped enable improvements in quality and patient satisfaction by reducing risk across transitions, supporting improved communication, and empowering care givers with the most effective recommended actions.

Tracking the ROI for cognitive machines requires a novel approach that accounts for the predictive analytic paradox and provides a clear perspective on how the solution is helping the organization meets its essential value-focused goals, including preventing adverse events and decreasing the instance of poor outcomes. This approach can be broken down into the following four steps.

Understanding the solution’s effective accuracy. The term accuracycan be used to hide the real effectiveness of a solution. For example, a model designed to predict an event that occurs 1 out of 100 times can be 99 percent accurate and miss the one event that it was designed to predict.

Instead of looking at accuracy alone, providers should understand the effectiveness of a cognitive machine. Effectiveness accounts for the ability of the machine to identify patients who are at risk of a condition or illness, the ability of clinicians to act on the recommendations and prioritizations, and the organization’s goals/program’s objectives. This measure combines the cognitive ability of the machine with the cognitive ability of caregivers to translate information into action.

Translate information into action. An organization’s ability to translate a cognitive machine’s outputs into results is a key metric in determining a solution’s value. For example, consider that a machine is firing predictions that are percent accurate in identifying patients at risk of a readmission. If the organization were to effectively apply the machine’s suggestions against each of the patients identified, 75 out of every 100 readmissions would be avoided. This is the “pure potential” of solution—i.e., the benefit that the solution would provide if the organization could apply interventions for every at-risk patient correctly predicted. But the reality is often different.

The ability to drive intervention effectiveness is influenced by the following factors:

  • Current capacity to perform the intervention
  • Ability to drive adoption of the predictive solution
  • Effectiveness of the intervention for a specific patient
  • Each patient’s willingness to engage with an intervention

Each of these points will affect the overall value that an organization can obtain from any type of prediction-focused solution. Capacity and adoption are likely to be optimized as the program enters a steady state. Interventions and engagement can be addressed through retrospective analysis and, in some cases, predictions. A solution that has built-in use cases and capability can help identify which interventions will be most effective and engender the greatest engagement by patient.

Assessing performance relative to short- and long-term goals. For most organizations, the number of adverse events will fluctuate, with many seeing a rise in number of occurrences that corresponds with a rise in the patient population. Short-term goals should emphasize a stop in the rise of events while chipping away at the overall number.

Long-term goals should focus on reaching the pure potential of the solution—acting on every correct prediction and preventing an adverse event from happening. Obviously, this goal remains an ideal to which every organization should aspire, but progress is possible, and it also is important for organizations to continuously track improvements.

Assessing the extent to which a solution is helping the organization meet its goals involves three key considerations:

  • Where the organization is today—i.e., the line representing how well the organization is applying interventions based on the solution’s cognitive capability
  • Where the organization might have been without predictive analytics—i.e., the line representing how many adverse events might have occurred without the solution, where the difference between this number and where the organization is today represents total avoided events
  • Where the organization could be—i.e., the line representing the potential impact that the solution could have with 100 percent intervention effectiveness

It is the second consideration that addresses the actual value delivered by the solution. 

Aligning to the Triple Aim. The Institute for Healthcare Improvement (IHI) developed the Triple Aim initiative to clearly state the goals of value-based care: to improve patient health, to improve population health, and to reduce the cost of care. When thinking about the value of a cognitive machine to an organization, it is important to outline targets and goals within the context of the Triple Aim. Doing so will provide a broader, more holistic view into the true value of the solution.

Aligning for Predictive Analytic Success

As with any new technology, healthcare leaders must drive adoption and engagement. This process starts with understanding the value of the solution, communicating it to the right people, and enabling the achievement of organizational goals through training, change management, and leadership engagement.

Once the right stakeholders have been identified and engaged, the rationale for the implementation and the anticipated benefits should be communicated, with expectations set that align to program goals. These goals should go beyond the number of patient lives improved to include improvements in the patient experience and goals to lower the costs of care. Performance should be monitored and messaged across the life span of the cognitive machine.

This approach is best realized through activities specific to each major stakeholder group (e.g., case managers, tele-health nurses, and inpatient care givers) and with leadership engaged in establishing improvement targets and encouraged to communicate with peers and the organization as a whole.

A project owner also should be identified to drive the project. This individual should understand how to politically navigate the organization, and should be provided with the tools and support needed to educate the organization about the cognitive machine solution, communicate the anticipated value, and manage adoption.

Buy-in from end users is essential. Key end-users can support communications and serve as administrators once the system is up and running. Training should go beyond the system to provide strategic context where the outcomes are about the patient and the cognitive machine is presented as a critical appliance to accomplish those outcomes.

In addition to engaging stakeholder and fostering organizational adoption, it is important to recruit the right talent. This may include data scientists, clinicians, data stewards/governors, and business analysts. Doing so can help an organization successfully establish and maintain a predictive solution regardless of whether it buys or builds that solution.

Next Steps

Cognitive machine solutions are becoming a core capability required for provider operations. Healthcare leaders looking to acquire such a solution should ask several fundamental questions to ensure there is alignment of these tools with organizational strategy:

  • Is the organization actively working toward value-based models of care and payment?
  • Is population health a priority?
  • Is the organization committed to using and applying these new solutions?
  • And if the assessment reveals the organization is behind in reaching strategic goals, is the organization prepared to make the investments that will help it maintain/improve its market position?

Answering these questions will help the organization assess whether a cognitive machine makes sense and where it is needed, and to build the business case to support adoption.

As cognitive machines evolve—with improved capabilities, more widespread adoption, and a heightened importance within the healthcare community—healthcare organizations will need refine the way they capture and define the value delivered by these solutions. This evolution is an intrinsic part of the greater shift to a healthcare system that rewards high quality over volume.

Ritesh Sharma is COO of Jvion, Johns Creek, Ga.

Leigh Williams is business systems administrator, UVA Health System, Charlottesville, Va.


a. Barrett, M., Wier, L., Jiang, H.J., and Steiner, C.A., “ All-Cause Readmissions by Payer and Age, 2009-2013,” Statistical Brief, The Healthcare Cost and Utilization Project and the Agency for Healthcare Research and Quality, December 2015; and Brem, H., et al., “ High Cost of Stage IV Pressure Ulcers,” American Journal of Surgery, October 2010.


googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text1' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text2' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text3' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text4' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text5' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text6' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-text7' ); } );
googletag.cmd.push( function () { googletag.display( 'hfma-gpt-leaderboard' ); } );