Welcome to identifying risk and segmenting populations: predictive analytics for population health. This is lecture c. In this lecture, we will continue to explore the domain of risk segmentation and predictive analytics in the population health context. We will offer a case study of how one tool the Johns Hopkins ACG system is applied within the Johns Hopkins Medical Center's own health plan and population health program. We will also talk about the future of the field, including the opportunities offered by the rapidly expanding electronic health records systems now found within most organizations. The objectives for this lecture are to discuss a case study of how a one common risk segmentation/case finding method has been applied to population health. Examine the role of various electronic data sources in risk identification/segmentation. Identify and discuss the developing frontiers in the population-based predictive modeling field. To begin, we'll examine a case study within the Johns Hopkins Academic Medical Center. The Johns Hopkins Healthcare Organization, JHHC, is a population based health plan, that receives a capitation rate for government and private organizations, to care for large groups on enrollees. This Managed Care Organization, MCO, uses the Johns Hopkins adjusted clinical groups, ACG system tool. In this lecture, we will offer a case study of how ACGs and other risk measures are use to help manage the care of the enrolled population for which JHC is responsible. This slide presents an overview of the Johns Hopkins healthcare's organizational structure. With further details regarding the population oriented care management services it provides. This chart depicts the flow of data for the cohortive individuals who are enrolled at JHHC It outlines the flow of data, taking claims data, lab and electronic medical record (EMR) data, and various ACG information, including the specific expanded diagnosis clusters (EDC), diagnostic components, and risk scores. Each of these inputs may be collected on a different timeframe, in some cases weekly, monthly, or annually. The organization has developed data warehouses for monitoring individuals and then sharing reports for various types of users. Users might be an individual case manager, a medical director who is monitoring the entire system or could be the patient's personal clinician at his patient centered medical home. Here is Johns Hopkins Healthcare's version of the pyramid. But in this case, it's in reference to people who have diabetes. As eluded to before, we are not just looking at that single condition, but also understanding the complexity of all conditions persons with this disease may have. JHHC constructs the various risk segments based on the ACG score, which reflects both the severity of the diabetes and the range and seriousness of other co morbidities. JHHC uses the patient's hemoglobin, A1c,derived from the EHR or lab data system. The clinicians among you understand that that's a very important marker for blood sugar level over a period of time. Based on this segmentation, the John Hopkins population management nursing team is able to select those persons in greatest need of outreach or other interventions. As shown from the previous Medicare analysis, the local situation is quite similar for the denominator population. Those patients who were capitated with the John Hopkins MCO. That is, the highest risk individuals use a large amount of services. Here, a total of $139 million of cost is expended by a relatively small group, 10,500 individuals. It shows the per member per month PMPM breakdown in costs of individuals at each level of the pyramid. So 543 individuals in this cohort who have diabetes and are also highly complex actually cost the system. $39 million over the year or over $5,500 per month per patient. Displayed here is an exampler output of detailed information on each high risk individual at John Hopkins Healthcare, MCO. This information is shared with the case manager nurse and represents typical output of the John Hopkins ACG software. It reminds us that once the large segments are identified, detailed risk information will be needed to try to understand the unique situation of each person in need. You may wish to review this sample quote, risk dashboard, unquote, carefully. It is an example of what can be derived with only claims data using a comprehensive risk adjustment tool. If EHR information is available the dashboard can even be more comprehensive. For example, including blood pressure and height, weight information. This report provides a structured starting point that the care manager can use to prioritize, plan and manage their assigned patient workload. The example is a real patient, a 60 year old male enrolled in a JHAC HMO. The probability that this person will be a high user next user is arrayed there. In this case, the likelihood is that this person has a greater than 50% probability of being in the highest group. These and other predictions, as explained in the previous lecture, as based on his current risk factor weight, based on large benchmark populations with similar characteristics. This also erased the various types of actual costs and current medical conditions. You can look at it slowly, it shows the presence or absence of different types of services and different types of coordination markers. In gaining an understanding of this person's needs and history, It is important to look at not only diagnosis but also patterns of care he received and how well coordinated it has been. This dashboard report shows the number of physicians seen. Those patients who have seen a lot of doctors are more at risk for inefficient care now we return to the pyramid, in this case, a three level pyramid. It reminds us of using reports and scores, as you've just seen. We can think about what a nurse case manager, or whole team. Team might do at a different levels of the pyramid. At the bottom level, there's outreach and education. For example, a smoking cessation program or weight loss program. The outreach could be via email or a web portal or mobile health app the second level might include remote monitoring. The program at Johns Hopkins is called TeleWatch. It offers telephonic prompts to participate in interventions such as blood glucose monitoring. It allows the patient to share feedback telephonically at the highest levels because you cannot on a nurse to every body in the 30000 person population, a nurse case manager might have a case load of 100 or 300. The care manager might be a social worker, an RN or a lay health worker from the community who is trained to work closely with the patient to address any needs he might have, such as housing or access to healthy food. The latter type of interventions are specially important for many Johns Hopkins patients given the hospital's location in an inner city community. To date, most predictive modeling and risk segmentation within the field of population health have used administrative risk data derived from hospital admissions records and health insurance claims information. Some also use surveys but the problem with survey is that there are many nonrespondents and it's very expensive to collect those data. The outcomes of most of the risk adjustment tools to date, have been on cost and utilization variables. This is a very exciting time in the field of predictive risk measurement. Many researchers are across the country, including at the Johns Hopkins Center for Population Health IT, are looking at all types of new sources of information, as shown here. That can increase both the accuracy and scope of the models. It can include the expanded clinical information found in the Electronic Health Records. It can also move into population and consumer information from a variety of data sources. We will discuss a few example of these new opportunities in the remaining part of this presentation. One new source of information that could be used to expand our knowledge of an individual's risk factors can come from the charting function of the Electronic Health Record, EHR. The core objective of the EHR is to capture information to support the interaction between the clinician and the patient. Some examples are, clinical findings, medical history, biometric information, blood pressure, body mass index, BMI, family history, symptomatology. One of the challenges is that, in addition to closed-end, quote, click, unquote, entries, the EHR is used in a major way to capture free text that is the doctor and nurse just taking notes as they did in the old days with paper and pen. And this does present some challenges, as most of the risk models derived from claims require specific fixed data fields. The clinical workflow may include clinical decision support, CDS, offering guidance to the provider. This workflow is useful for documenting who knew what, when. It can be use particularly for predictive modeling of a clinical phenomenon, to look at the flow of the symptomatology. For example, when did a lab value changed and when did the doctor respond to that information? Provider order entry known as POE or CPOE such as e-prescribing or test-ordering can also be used to drive risk information. With such data, it’s possible to examine the prescription that doctor ordered versus what medication the patient actually obtained in the pharmacy. Most of the predictive modelling tools to date, use pharmacy claims information, not the e-prescribing. Understanding the difference between what was prescribed and what was actually obtained is an example of some information that can be mind from the workflow function of the EHR. Increasingly, results investigations which generally include both laboratory and imaging, X-ray, EKG, electrocardiogram and other cardiovascular is all digital. One of the disadvantages of insurance claims information is that, it does not have the clinical texture, such as this information. From claims, you only know that a test was ordered, you do not know the result. Inputting these factors into the predictive modeling process would potentially represent a great advance. The domain of EKGs and imaging using so called PACs, picture archiving and communication system and other computerized representation of the images is increasingly coming online. But, these are sometimes very difficult to input into a risk adjustment model, as the images are not standardized. Labs on the other hand, report very specific results. And, one knows whether it is in or out of the normal range. These are rapidly being integrated into predictive modelling tools. We are using many new types of devices. Home devices, telemedicine, mHealth, quote,mobile health, unquote, sensors, Fitbits. Apple is trying to market its iWatch to healthcare systems. The idea is to integrate consumer movement, heart rate and other things captured on the sensors of the iWatch directly into the EMR. That's a little bit in the future but this increasingly will be part of predictive models. Certainly for a long time consumer information from health risk appraisal, HRA, surveys have been input. Again, the problem is that sometimes, sick patients don't complete those surveys and the response rates are fairly low. But obviously, consumer information and preferences and functional status, how well they're able to accomplish various tasks, is also an important data source. And to the extent that consumer-based eHealth facilitates capturing these data, this represents a potential advance. Also on the consumer side, some commercial insurance companies and others are now mining the same financial information and internet information that other marketers are using. What type of magazines you read? What type of websites you visit? What is your credit rating? What community you live in? And these things could potentially be used not just to sell you something, but also to understand your risk level and improve your health function. Understanding the patterns of electronic interaction between consumers and doctors will gain an importance for those assessing risk. This type of interaction is increasing rapidly in wired organizations who use email and web portals. Also related to this is interaction not between providers and consumers but between consumers and other consumers via social networks. Some studies have suggested that, understanding these patterns can also help understand current and future health patterns. The area of community surveillance and public health data is one that's increasingly being blended into medical care delivery system. It still relatively new to include geographic level data into a health plan or a delivery systems predictive modeling score for case finding but this is beginning to happen. For example, if a person lives in an area with high environmental risk, low food availability or a high rate of communicable disease, these are factors that might be included in calculating their overall risk score. Up until now and in most of this presentation, predictive models and risk adjustment have focused mainly on healthcare costs and utilization because those are the types of data that have been available. As EHR data are coming online, increasingly, they are becoming interoperable, that is, linked across providers. This will allow us to capture patient outcomes across the entire care continuum. As this happens, we can use EHR and other data sources to identify new endpoints of population-based predictive models and risk measurement systems. That's exciting. Here's a summary of some of the end points or targets. These new models may predict the trajectory of disease over time. For example, when is it likely to get worse, or better? They might predict the health in a community, following more the public health model of surveillance, the model might target functional status. For elders, it's particularly important not to just look at disease, but also function. There could be people with significant disease who are high functioning. And there can be people with lower levels of disease who are low functioning. It's important to understand biometric attributes, blood pressure for example, or blood sugar. There have been great advances within clinical forecasting and modeling in some cases using genomic information. To date that has mainly been of scientific inquiry or focused clinical conditions. This area of work hasn't been expanded into the population domain yet, where we attempt to understand risk in communities or other large denominators. But in the future, it certainly will be. Continuing on with new targets for predictive models are social needs and challenges. One could predict, for example, in a disadvantaged pediatric population those who didn't have adequate access to nutrition, things that weren't there own medical outcomes, but really are all important to health. There are models and approaches that are starting to do that. Also knowing how important certain consumer behaviors are for health, take exercise or smoking as examples, it would be helpful to have models that predict a consumer's tendency toward high risk behaviors. And while there have been models from a research perspective looking at longevity and mortality, in the future these will be integrated more into a population based perspective. Today some models are being applied to help guide end of life and hospice care programs. This map depicts an activity going on in Baltimore, Maryland. Where the Health Information Exchange known as Chesapeake Regional Information System for our Patients or CRISP for short is working with Johns Hopkins and other organizations to try to do predictive modeling at the entire state level. Linking in data from all of the hospitals. Potentially data from all of the health insurance plans and increasingly data from the electronic health records to try to identify who is at risk for readmission. There is a unique situation in Maryland in that the hospitals are paid a flat global budget. And they are being held accountable for an array of outcomes, including readmission. So there is a lot of incentive to try to predict and intervene at the population level and not just wait until the patient is admitted. This field of risk segmentation and predictive modeling in population health is evolving very rapidly. As new electronic data sources come online, and as our health care system increasingly reorients itself towards population health. Before we close this lecture and unit, it is worth reviewing some of the major ways health IT is likely to lead to advances in the coming years. The old model of medical care, just focusing on the patient in the hospital bed or in the ambulatory office, has given way to, really, an understanding of the cohort from which that patient has come from. As well as understanding how we can be most efficient at targeting people with great need. Increasingly with new data sources and analytic tools we will be able to identify who was part of the population denominator and who will fall into the numerator of the various risk pyramids we have showed you. This sounds simple, but it isn't. Speaking of increased denominator focus in the past, virtually all predictive modeling and case finding focused on those cared for by a provider organization or insurance plan. Now with community-wide interoperable medical care data and non-medical public health and social data, it will be possible to accomplish this process for geographic communities. Most of the data sources today have been for risk calculation and have come from fixed, structured data sources. Increasingly through data science techniques, we'll be able to do data mining that will identify unstructured data. The classic one could be the doctors or nurses notes that are just in text form. But there's a lot of other unstructured data that can be captured. Historically the risk information and predictive modelling have focused on outcomes that are based on cost or a medical model. But as we try to understand the multifaceted equation of what a population values, it could be social, it could be function. Increasingly it will be possible to pull together a broader array of information and develop a dash board for the care manager, ultimately for the consumer, for the clinician of the different variables, the risk, and what can be done to intercede. Just as there are been great advances in clinical decision support, there are are likely to be great advances in population decision support, tools using predictive models, risk segmentation that will help population management really enter a new era at the patient panel level, at the community level, at the system level. This concludes Lecture C, Identifying Risks and Segmenting Populations Predictive Analytics for Population Health. In this lecture, we reviewed a case study of how a population health oriented organization, the Johns Hopkins Healthcare Managed Care Organization implemented risk segmentation to better manage the large populations of persons for whom it is responsible. We also offered some incites into new uses of electronic health records, and other data sources that will likely expend this field considerately over the coming years. Let's review the main points of this unit. Risk adjustment and predictive modeling are essential for managing risk in today's healthcare system. There are several methodologies for gathering electronic health information on patients so that we can segment populations into sub-groups to identify those with higher-risk levels. Predictive modelling tools are developed using large benchmark populations. Both analytic approaches and clinical logic are applied to these tools. We looked at a commonly used predictive modelling tool, the Johns Hopkins ACG System, to learn how these tools are constructed and used. And lastly, EHRs and other new sources of data will guide future developments in predictive modeling tools.
