Welcome to Identifying Risk and Segmenting Populations: Predictive Analytics for Population Health. This is lecture a, in this unit, we'll provide an introduction to commonly used case identification and predictive modeling tools. There are a variety of tools out there. The team at the Johns Hopkins University who designed this unit are the developers of one of these tools that is widely used across the world. So we will use that as an example. However, there are many other widely used tools as well, and they're all fairly similar. The objectives for this lecture are to define and discuss perspectives related to the concept of risk measurement and segmentation within the population health context. Describe the commonly used case identification/predictive measurement/modeling tools. In case you have not been previously introduced to the field of population health, our working definition is one of encompassing healthcare, communities, and public health agencies. It takes a perspective that includes both communities and denominator populations defined by geography. But also groups of individuals who may be defined by participating in a health insurance plan, or being cared for by a specific accountable care organization, or integrated delivery system. There are several working definitions that will be helpful as we continue in this presentation. Let's review them briefly. Health information technology is the application of electronic health records, EHRs, sometimes also called electronic medical records, EMRs, and other types of information technology, sometimes called e-health. There are a lot of other names for these technologies, that can be used for the delivery and management of healthcare. And when we talk about healthcare, we do mean both medical care delivery, public health and other types of services that improve the health of individuals and populations. Population health informatics is the systematic application of health IT, and other digital technologies as well as information science, sometimes called informatics or data science. There are various terms that are applied, but the ultimate goal of population health informatics is the improvement of the health and well-being, not just the medical care of a defined community, such as a city, or a state, or a town, or some other target population. The term population health is used widely in the area of chronic care management, it is also applied in community health. It is a term that crosses over between medical care and public health, and again, public health is often defined as government health programs targeted at a jurisdiction. Nowadays, we use the term analytics, or big data, quite often. We offer this definition of healthcare analytics that is leveraging electronically available healthcare data, perhaps medical care, perhaps public health, to enable actions that improve the effectiveness, efficiency, and equity of health systems. Efficiency, the ratio between cost and value, implies that we pay attention to not just what the outcome is but also how much it costs to achieve that outcome. And equity means ensuring that all members of the population have equal access to care. Terminology in this domain can be confusing, but there is a general consensus of the terms predictive modeling and risk adjustment. Risk adjustment means taking health status and risk into consideration for healthcare finance, payment, provider performance assessment, and patient outcome monitoring. Often in this context, the term for health status is risk. Within the insurance context, but outside of health insurance, the term risk may refer to financial risk. But here we're talking about a clinical risk, a health risk. So here, the term risk adjustment means taking a population's health status into consideration when undertaking an administrative activity such as a healthcare provider payment. In other words, if you're setting a provider capitation rate, the provider would receive a higher budget if this population consists disproportionately of sicker people rather than healthier people. You would also want to take this type of risk into consideration when assessing the efficiency of a certain provider. For example, how many of her patient should receive a certain test. What is the overall pharmaceutical budget for her patients, sometimes this type of analysis is called provider performance profiling. And for sure, if one is looking at the outcome of a provider, an outcome of health for a community, one wants to know whether or not the patients being served started at low risk or high risk. Otherwise, such comparisons across two populations would not be fair. So, while the term risk adjustment is also used in epidemiology, statistics, and insurance, here we're using a population context to look at the characteristics generally associated with a person's disease and morbidity. And adjusting for that when one is looking at various outcomes, be they financial or clinical. Closely related and increasingly important, though sometimes overused, is predictive modeling. Although it is a relatively new term, the concept has been around for a while. In general use, it has statistical forecasting or economic connotations, but within healthcare it's become widely used, particularly in the field of population management. In regard to predictive modelling, we offer this definition. The perspective, that is in the future. Or concurrent, that is in a more recent period of time. Application of risk measures, that is health status measures and measures we'll talk about in this module. And statistical technique, could be a forecasting technique, could be a computer science technique to identify, quote, high risk, unquote, individuals who would likely benefit from care management interventions. And we use the term care management fairly widely. So, predictive modeling, in the population health context, makes use of available risk data and statistical analysis to array a population based on our understanding of likely outcomes. Once one has been predicted who is going to be at risk, let's say for hospitalization, or a bad outcome such as a death or increased disease. We array our population into what is frequently termed a population risk pyramid. In this pyramid, we array all individuals in a cohort from low to high in terms of the predicted outcome of interest. We then use this stratification to allow us to select individuals from within the population who are at greatest risk, and who are likely to have the most need for a certain type of service or intervention. To review. Although the terms are closely related and they're used in a variety of ways, risk adjustment really focuses more on taking into consideration risk variation when assessing an outcome. And predictive modeling usually is intended as a mechanism to identify high risk cases for some type of intervention. Throughout this presentation, we'll use various pyramids. These are population pyramids that array everyone in a community. Everyone a target population, say within an accountable care organization, ACO, integrated delivery system, into one or more risk levels. Now let's talk about how one does this. How does one identify risk? How does one segment the population? That is, put people at the different levels of the pyramid. It'll raise a population from low to high. Low risk, moderate risk, high risk, advanced disease, or multiple diseases. Again, for most people, when assessing total health status it's usually not a question about having only one disease with high severity. Most people who have one disease usually have multiple diseases. So really, high risk is a combination of both of those phenomenon. Not to mention all of the social and non-medical risk. We'll talk about that in a moment. The pyramid shown here, the original of which was developed for the National Health Service in the UK. Also presents and reminds us, that people at all risk levels can get care in three different ways. They can care for themselves or in the community. It could be behavioral, getting exercise. It could be making sure they do activities at home. They can get regular clinical care, seeing doctors, nurses, interventions that are fairly conventional. Or they can be cared for in a more intensive way through special services, such as a nurse case manager or home visitor. That's usually considered the top of the pyramid. Although different individuals at various places on the pyramid, as you can see from the graphic, will benefit from all three levels of care. This is actually a simplified graphic. But the focus here is, how does one take this theoretical presentation, and actually translate it, and actually translate it with data to identify specific individuals? This graphic is a unified framework for health care, prevention, and disease management. This framework has three dimensions. The stages of disease, the levels of population health intervention, and the continuum of prevention. Understanding risk is essential at all stages of this framework. Look at the green boxes at the core of the framework. We present the continuum of individual health. A person, a patient, a consumer, could potentially be at risk. As you can see all the way to the left of the framework. Some of us are at higher risk or lower risk of various diseases, based on where we live or what conditions we may have, and of course, our genetic predisposition. Although in most cases, that's not yet measurable on a wide basis. Moving from the core's left to just one box to the right, a person can increase in risk, and get early symptoms of disease. Let's say, for example, someone is at risk for high blood pressure, and their blood pressure is slightly elevated. Or someone is at risk for diabetes, and their blood sugar is slightly elevated. In that instance, the person would be going from at risk to having early symptoms. Moving even further to the right in the framework, a person can actually have the disease. And then eventually the disease, say hypertension or diabetes for example, could lead to negative outcomes. Chronic care management spends a lot of time managing risk in an organized structured basis. One needs risk measurement and case-finding tools to accomplish this. Let's turn our attention to the blue boxes at the top. Community interventions, clinician/patient interventions, and delivery system interventions. For example, if one has a population of interest that may be at risk for diabetes or hypertension, exercise and good diet at the community level are important. Obviously a clinician can provide preventive care. Good supportive value based care, with medication and guidelines. Again, that's really more viewed as clinical quality improvement, rather than population health. But an entire delivery system, whether or not it's a patient-centered medical home, PCMH, or accountable care organization, ACO, really concentrates on the big picture. And ensures that the system attempts to address all individuals on this continuum, not just those off to the right-hand side of the framework who are sick. And one needs risk assessment for all of this. On the bottom of the framework, we offer a reminder for those of you interested in classic prevention, be it clinical or public health, that we really have several levels of prevention. One is community-based prevention. The other is primary disease prevention. And then secondary prevention. A term that's sometimes used to reflect individuals who already have the disease. But we're trying to avoid a severe outcome of that disease. So this is a paradigm, not the only paradigm that is useful. Not just for population health, but in understanding how we can use case finding and segmentation to identify people and support all of these ranges of activity. It's important to remember that this paradigm is useful for population health, as well as case finding in segmentation, to identify people and support all these ranges of health and disease. We share this framework as a way of quote, opening the black box, unquote, a bit to give you a sense for all the ways that predictive modeling and risk measurement may be useful within the population health context. The next series of slides represents the various types of care management, clinical management, and health care management activities that require risk segmentation, case finding, or predictive modelling. Again, these are all interrelated terms. Here we offer another pyramid. This is one that we developed at the Johns Hopkins University Center for Population Health IT. CPHIT, or CPHIT for short, which is also the research and development home of the Johns Hopkins ACG Predictive Modeling System. We'll talk more about that soon. The Johns Hopkins ACG system, is one of the two or three most widely used population based, risk adjustment predictive modeling tools, in use across the United States. As well as in about 20 nations. Again, it is not the only system. There are several out there, such as HCCs, CDPSs, CRGs, that will be discussed later. But because we are very familiar with the ACG predictive modelling system, we're using it as a model that is pretty representative of the field. Looking at this pyramid, you'll see that it identifies the different segments of a typical population, that is everybody. Whether or not they come in contact with a medical care provider. This is a very important point in population health. And why it is distinct from clinical medicine, is that it will include everybody, whether they knock on the doctor or nurses door or not. So the bottom of the pyramid, in this case, is defined by whether the segment of the population are users of health care or non-users of health care. And each successive layer of the pyramid, by the way, includes everyone above it. The segment of users and non-users will include everybody. The next run up on the pyramid includes those people who use healthcare, whether or not they are low risk or high risk. The next run up on the pyramid, recalling our previous framework graphic, includes individuals who may have modest levels of risk such as a single high impact disease, or several lower impact conditions. And then at the tip of pyramid, there are those people with multiple high impact morbidities, which lead to an overall higher risk burden. The pyramid presents a way of thinking and a way of arraying the entire population, either in a community or in an integrated delivery system or health maintenance organization, HMO. Off to the right, let's look at several administrative and population management applications, represented as columns. That may be applied to the various population sub-groups with different risk signatures. The column all the way to the right of the slide reflects activities that need to happen for 100% of the population. For example, if you're in an HMO that has 100,000 people who rely on you for care, you're paid on a capitation basis for all of them. So for each person, you get payment whether or not the come in to get care. You are responsible for those individuals, users and non-users alike. Similarly, for all of those persons, whether or not they get care, you have to monitor quality. You may have individuals with specific diseases for which you know they were treated two or three years ago, who don't come in, even though they should. In population health, in accountable health, you are responsible for that individual. And likewise, it is important, as we showed earlier in this framework, to really understand individuals at risk, and reach out to them to assess their needs before a condition can advance. The next box, practice resource management, is applied to the top three segments of the pyramid, basically all users. You need to understand how to provide care efficiently to this cohort. These tasks would include performance profiling. In the two higher levels of the pyramid, is where care management or disease management comes into play. Population health management is now sometimes a term that's used synonymously with disease management. Much of the population management today focuses on the management of chronic disease. The top level intervention is usually some sort of case management that might involve assigning a nurse to an individual at very high risk, to help coordinate the care process. Different labels may be used for this, and other types of intervention. The main point is that for different parts of the population at different levels of risk, there may be different management applications for the risk adjustment or risk-based case finding. Here, some of these applications are broken down into a little more detail. They describe the array of applications that generally fall under the category of population health. Though some might cross over into health insurance or healthcare functions, that focus on individual patients or services, and not just on populations. First there are financial concerns. In the US and in other countries, there are salaried organization, budgeted organizations, and capitated organizations that are paid up front. And if one is going to care for a population that is more ill than another population, or at higher social risk than another population, one needs to identify and segment the risk. Likewise, one must decide how many services are required while trying to stay within the allocated budget. Examples of services could be hospital care, specialist consultant care, pharmaceuticals. From a financial perspective, one needs to understand the risk. One also needs to understand efficiency profiling, which as noted is often termed, performance profiling. Understanding cost efficiency and outcomes and process quality markers is central to the accountable care organization. When payment is linked to achieving performance targets, one also needs to take risk into consideration. It wouldn't be fair to look at outcomes, let's say hospitalization or how advanced certain diseases have become within an enrolled population group and compare it to another group, without taking risk and health, or risk status into consideration. For quality improvement analysis, or ongoing monitoring, one also needs to take risk into consideration. Particularly for outcome measures, but also for process measures too. As described previously, there are a variety of phases in the care process where risk segmentation tools can and should be applied. To ensure that the right people are identified for inclusion in various population health management programs. Especially when one is extending the care management to a geographic community, it is essential to also consider socioeconomic factors, in addition to clinical or healthcare factors. When applying these tools to assess community need or to identify those who would most benefit from targeted interventions. These tools also have many applications for evaluation and research. One cannot assess whether a program or intervention is working, unless one controls for risk across the cohorts undergoing the intervention, and those in the control or comparison groups. So you can see, there really are many applications for digitally based risk adjustment, risk segmentation, and predictive modeling. It's all digital. No one does this manually. We now use digital information to segment risk or identify high health status groups and low health status groups, who need different levels of attention. This is an analysis of chronic care co-morbidities across the US Medicare population over the age of 65. It gives a pretty good sense of why risk measurement or risk segmentation is essential when understanding or assessing care delivered to a population group. It also shows the importance of understanding and considering co-morbidities. For this table, we are looking at chronic condition co-morbidities. That is cases where person has two or more medical conditions, which may or may not be related. At the same time, this was an analysis done using US Medicare insurance claims data by faculty at the Johns Hopkins University. When one talks about using digital data for predictive modeling and risk adjustment by far, the preponderance of activities to date have been with health insurance claims data like these, but electronic health records are increasingly being used as a new data source for this purpose. In this table, we've arrayed the entire population of elders into levels that reflect how many common chronic conditions they have. You'll see looking from the bottom up that 53% of the individuals have either zero to two conditions. In most cases, they have one or two of the in-scope conditions considered. 27% have three to four conditions. And a fairly large percentage, 20% have five or more conditions. The table shows the average amount of medical care services and related cost for individuals within each subgroup based on their chronic condition count. Although this approach to segmentation is far simpler than the approach that the most commonly used risk adjustment tools use to categorize people, it gives you a good sense of how this process works and how important it is to understand the individuals clinical characteristics when assessing patterns of care, need or resource use. Take your time looking at this table. Some of the patterns are fairly profound and really gives you a sense of the importance of risk segmentation. You will note that the top 20% of individuals account for 66% of Medicare's cost. And on average, this top group sees 13.8 different doctors during the year and they fill 49 separate pharmacy prescriptions. This simple analysis shows why we should all be very interested in risk segmentation when it comes both to individuals and populations. That is both to help a patient in need and to move our healthcare system towards efficiency, and effectiveness at the delivery system population or community level. There are a variety of widely used well-validated methodologies that are used in different contexts. They are often labeled with various names depending on the context, risk adjusters, risk measurement tools, predictive modeling tools. We will discuss several of them. Most of them got their start in the domain of healthcare utilization and they were calibrated to try to understand or explain healthcare costs, and that usually includes all costs or sometimes just hospital costs. Some models were developed to try to identify those persons likely to be hospitalized or rehospitalized. Today, most models, regardless of their original focus are now used for that purpose. Often, these methodologies are developed, so the statistical models that they are based on are calibrated to identify the tip of the pyramid. Say, persons in a population, whose overall costs are within the top 1% or top 5% of the distribution. Some models have variants that are recalibrated to identify different types of costs, say only pharmacy costs or to focus not just on cost, but future events. In early days, there was a concern that some insurance companies would use these risk adjustment tools to try to identify expected future healthcare costs to charge higher rates for sick people or to not enroll them at all. That is no longer allowed since the Accountable Care Act, ACA has been put in place. When setting future payment rates, say for an HMO or other health plan, it is common to use this year's health status, say as measured by diagnoses codes to predict next year's costs. That is one reason the term predictive modeling is used. Although as discussed earlier, typically that term is used for case identification applications within care management programs and risk adjustment is used for financial applications. Though the term prediction should be used loosely, because sometimes the methodology targets a future time period such as next year, but one can also calculate a risk score to understand patterns of use within a current year. In that situation, it's not exactly a prediction. It's really more of an association. For example, this would be the context such as we did previously with the Medicare table when we used diagnostic characteristics to help stratify patterns of use for this year. Moreover, if one wanted to send in a nurse to help someone out who is really having problems and has need here and now, one would try to focus on the concurrent risk of characteristics of the person rather than their characteristics in some past time period. So again, the target, the end point, the outcome of these various models can be different time frames for different applications and the type of service targeted may be different. One widely used method is what is termed Hierarchical Condition Categories or HCCs. The system was originally developed at Boston University. A version of it in the public domain is used by the Center for Medicare and Medicaid Services, CMS that runs the Medicare program. There is also a version that CMS applies for qualified health plans in the ACA's marketplaces. A commercial version is also distributed by a company called Verisk. HCCs are also used to adjust the performance targets for Medicare's ACOs. Their strength is financial applications, because they are calibrated for that task in particular. On the other hand, the developers of HCCs have a full array of tools on a commercial basis that are available for all types of care management activities. Another widely used methodology developed at the Johns Hopkins University is officially called Adjusted Clinical Groups or ACGs. They were originally called ambulatory care groups, because they were meant as an adjunct to the inpatient Diagnosis-Related Group or DRG system. These are widely used by many Medicaid programs and for financial application. They're used by many private health plans and in government agencies for care management. In addition to the US, they're used in 20 or so other nations. Chronic Illness and Disability Payment System, CDPS is a methodology that was originally developed for Medicaid programs. Clinical Risk Grouping software, CRGS is a methodology developed by the 3M Corporation who also developed the DRGs. It is used in various Medicaid programs and private organizations in the US, and various other locations globally. A widely used public domain method often used in research is the Charlson Index. It counts a relatively small number of chronic conditions, but it's a useful standardized tool. The federal agency for healthcare research and quality, AHRQ, developed a risk adjustment tool called Clinical Classification Software or CCS. It is in the public domain. It's used often by researchers It counts various ICD codes. Again, ICD stands for the International Classification of Disease and it is the standardized method to used to categorize individual diagnoses. In addition to those general tools that are based on administrative data, there are also a series of tools that are derived from a survey or an assessment completed by a nurse or doctor. One commonly used tool that can use surveys and administrative data to identify persons at risk for hospitalization or death is called the LACE method. Each letter in LACE stand for a component of the measure. The L stands for hospital length of stay. The A stands for whether past admissions had high acuity. C stands for comorbidities, and E stands for whether or not the patient had to visit the emergency department. It is used by some health systems to identify persons at risk for readmission. Within the general field of medical predictive modeling there are many other tools that focus not on populations but rather on sub groups with specific diseases, often treated in the in patient environment. For example, which patient in an ICU is at risk for a bad outcome? Many of these tools involve data collection that are clinical and not driven by today's commonly used insurance claims data systems. We will not discuss this class of tools today the methods we highlight are some of the most commonly used tools but there are other methodologies used both in the financial risk adjustment domain and also the population based case finding domain. You can learn more about each of the tools we highlight as well as others with a simple online search. At the bottom of the slide, we offer a reference of an analysis supported by the Robert Wood Johnson Foundation that compares and contrast various methods and the challenges they and the field face. This concludes lecture a, identifying risk and segmenting populations, predictive analytics for population health. In this lecture, we saw that in today's health care system, risk adjustment and predictive modeling of the population is essential for many applications. Using computerized information describing the individual, these methodologies are used to segment the population into sub-groups to identify those with higher risk levels.
