Welcome to Population Health, Unit 7, Population Health Management Interventions. This is lecture a. The learning objectives for Lecture a are to, one, describe the population health data necessary for segmenting the population into risk cohorts. Two, differentiate the key cohorts of a population by the degree of risk. Three, analyze the root causes of risk in a population by utilizing certain socioeconomic, behavioral, electronic medical record data, and other demographic data. Four, explain the processes and key decision points by which interventions are prioritized for segments of the population. The most well-cited definition of population health comes from an American Journal of Public Health article. Quote, was is population health, end quote, by Kindig and Stoddart, who define population health as the health outcomes of a group of individuals including the distribution of such outcomes within the group. In addition, there are definitions that are more specific. Dunn and Hayes' definition, in their Canadian Journal of Population Health article, quote, toward a lexicon of population health, end quote, includes influences on health as part of the population. Quote, the health of a population as measured by health status indicators and influenced by social, economic and physical environments, personal practices, individual capacity and coping skills, human biology, early childhood development, and health services focuses on the interrelated conditions and factors that influence the health of populations over the life course, identify systematic variations in their patterns of occurrence and applies the resulting knowledge to develop and implement policies and actions to improve the health and well being of those populations, end quote. In 2012, the institute for healthcare improvement, IHI, published the population health model. It was adapted for the Affordable Care Act by the Academy Health Organization and is represented here as a framework of interventions on the bottom and factors that lead to effective interventions. In determining health interventions for population health, we consider equity as a major contributing factor in the following two domains, one, prevention and health promotion, and two, medical care. Prevention and health promotion are influenced by upstream factors, which include socioeconomic factors and the physical environment. There are also individual factors, such as a person's genetic endowment, their spirituality, behavioral factors, and physiologic factors that determine resilience at an individual level. In terms of medical care, there are intermediate outcomes that are related to disease and injury. The upstream factors and the individual factors will effect intermediate outcome and also affect states of health. That is, people's ability for health and function, as well as their mortality. These then lead to quality of life or well being. Johns Hopkins HealthCare in Baltimore has several population health definitions, which are gleaned from population health definitions mentioned on previous slides. Population health is a cohesive, integrated and comprehensive approach to health care that considers the distribution of health outcomes in the population, the health determinants that influence the distribution of care, and the policies and interventions that are impacted by the determinants. Population health management then is a process. It is the process of addressing population health needs and controlling problems at the population level. So there are strategies that address population health needs. Johns Hopkins HealthCare has developed a population health conceptual model, which begins with population prioritization and ends with evaluation of population health management interventions. We will discuss their health conceptual model in pieces. And we begin with the population assessment at the top of the model, which is highlighted here and is also enlarged on the left hand side of the slide. This portion of the model involves using all available data to understand the morbidity of a population. The health priorities, health risk, and the target for intervention. The population health assessment is often informed by claims data, administrative data, demographic data, and other data that can be gleaned from electronic medical record clinical data. That population assessment is pulled into a population database. Then, an analytics team interprets and prioritizes the health needs of that population. What is, for example, the occurrence and prevalence of conditions, chronic conditions, and how does that information lead to informing a population health strategy? Risk stratification and segmenting also happens at this part of the conceptual model. This risk stratification is done according to the needs of the population. It is necessary to have strong analytic teams as well as an IT infrastructure and IT support to aggregate data from multiple data sources. Below the population assessment box in the conceptual model and enlarged on the left hand side of the slide are the upstream factors and individual factors that affect the ability of a population to realize better health and better quality of life. These factors inform the development of primary strategies and interventions for the population. In developing these strategies, it's necessary to consider the social needs, mental health needs, physical needs, and lifestyle and behavioral needs of the population as well as the environmental factors, social factors, and genetic and biologic factors, which can profoundly influence disease and injury within a population. The population health process, then, is a very data driven one. In terms of population health management support, data are critical for determining interventions and the needs of a population. This table helps to further iterate the population health process. First, data are collected from multiple sources, that data must be then integrated and cleansed. And there needs to be ongoing management strategy for the integration of further data on a monthly or daily basis. Once data have been archived and updated, forming an analysis ready data set, the population health processes described on the previous slide can begin. Those population health processes begin with a population health assessment and with the development of condition registries, which may include conditions that may indicate that a person has a chronic medical condition, may be at risk of falls, had a special health need, or has functional limitations that affect activities of daily living. After the assessment, comes focus. We focus our analysis and our summary on important health status indicators, then we begin to predict. We develop and test tailored risk prediction models. We implement prediction models and we try our best to predict who, in our population, is at risk for higher utilization in the future, higher cost in the future or who is at risk of developing chronic disease. Once we have developed those predict models, we move to program development. And in that phase, we use ad hoc reporting to assist us in program development. We have outcomes and evaluation planning, and we have measure development and definitions. So what are our metrics and how will we define those? And how will we develop those from the datasets that we have access to? We identify and stratify individuals. So we get down to the specific patient level. We implement our identification algorithms. We identify gaps in care. For example, we may look at whether they are diabetics, who have not had hemoglobin a1c measured in the past year. We also have patient list with pertinent demographic and risk factor information. Then once we have identified and stratified the population based on risk, then we can begin our management and outcomes reporting on monitoring data boards, so that we can monitor our interventions over time and apply continuous quality improvement methods. Identification and application of benchmarks is also important. So how does our population compare? After risk adjustment, after adjustment for age, and demographics and other indicators, such as complexity of disease, to other populations, and other benchmarks. It's also good to provide information to physicians and provider groups. So after risk adjustment, how does their population appear to be functioning? Or appear to be utilizing services based on their profile? Finally, evaluation is important. What are the effects and costs savings of the intervention? It's important to disseminate your findings through written and oral reporting. Tailored to both technical and non-technical audiences. This table provides a review of data sources for measuring population health. In the column headers are the sources. EMRs, self-reported risk assessments, medical billing, pharmacy billing, labs, and monitoring devices. The types of data that can be gleaned from these sources comprise the row headers, which can be subdivided into four general categories. Patient demographic information and patient reported health perception and behaviors, shown on this slide. In patient health outcomes and utilization and cost types, shown on the next. The electronic medical record provides much data regarding patient demographic information, past medical history, and family medical history. EMRs may also include patient reported health perceptions and behaviors. So we can ask patients, for example, about their knowledge of a disease. In a ten minute test of self-knowledge and self-management we can ask patients about their symptoms, about their functional status, exercise habits and nutritional intake. We can interview patients or give them touchpads to gather this data, but often that is electronic medical record data and patient reported health data is a result of a face-to-face encounter with a physician or other healthcare provider. Many employers and many health plans ask their employees or members to take an annual health risk assessment, and the same data sources can be obtained regarding perceptions about behavior and perceived health status, etc. Note that the medical and pharmacy billing data don't provide all the population health data points that can be collected from patient-reported health behaviors. So combining electronic medical record data with claims and billing data is important. At times, we have patients using wearable devices or using some sort of remote patient monitoring, though it's not usually possible to get past family medical history or past medical history from these devices. It is possible if incorporated into the device, to get patient reported health perceptions and behaviors. And we can incorporate that into the set of data that we're collecting. On this slide, appears a continuation of our data sources table. If we continue with electronic medical records, we find that we can get prevalence of illness and comorbidity. And we usually get lab results, measurements of body mass index and blood pressure. Medication adherence though is not usually included in an electronic medical record, unless it's specifically measured by a provider. Results from quality of life surveys can be found on EMRs. Now what we cannot get from electronic medical record data, generally, is information related to hospital admissions, lengths of stay, re-admissions or emergency department costs and utilization. We can generally get primary care and specialty care data, lab and radiology utilization, though often, not cost. Prescribed medications, yes, but not medications that have been filled by the patient. Dialysis visits, post-acute care, hospice services, and vision and dental services are often not recorded in an electronic medical record. On the self-reported health risk assessment, that is the annual assessment that an employer or a health plan may ask its employees or members to take. We can get patients to self identify prevalence of illness, usually patients are not asked about lab results. These assessments can ask patients about their body mass index and blood pressure, but we have to question the accuracy of self-reported data. Utilization and cost types are not generally accurate in self reported health risk assessments. On medical billing data however, we can have most of the cost and utilization types identified because facilities are billing for those services. So we will have ICD-10 diagnoses that allow us to capture prevalence of illness and comorbidity. Generally, we'll not have on medical billing and pharmacy billing data, lab results, BMI, blood pressure, etc. What we can get additionally from pharmacy billing data is data on medication adherence. This is usually measured as what we called a quote, medication possession rate, end quote. So if a medication was prescribed to be taken once a day for 60 days, but it took the patient 120 days to refill it, then we can say that the patient adherence to that medication regimen was only about 50%. Generally though, in pharmacy billing data, we don't have information on hospital emergency department visits, other sites of care, or facility and service information. Lab data is often inclusive of illness and comorbidity prevalence. But many times lab data will have a rule out diagnosis, such as rule out diabetes. So we don't want to capture those as definitive diagnosis on lab data. Device monitoring, again, it's possible to get patient health outcomes data. Possible to get medications prescribed and filled by the patients in terms of self-reporting, but not likely to get cost and utilization data. So when the past two slides, you're getting a sense of the sources of data and how each source of data is valuable, but not 100% complete. Johns Hopkins Healthcare gets population data internally from claims, from for example, their Medicaid population, and they get their laboratory data. Johns Hopkins Healthcare has their intervention tracking systems, so their population health interventionists are recording and measuring data internal to Johns Hopkins, and they have the EMRs from providers. And then, they get Medicare data externally from the centers for Medicare and Medicaid services. They also have external vendor data such as from mental health providers, all of these data are put in to there population health database. They have a process of cleaning, manipulation and integration. So they produce analysis ready datasets and they produce population and variable definitions. Both of those are combined and the population and variable definitions are applied to their datasets. These steps result in the creation of a statistical analysis presentation. Now the outputs from that analysis presentation are many and multiple for multiple uses. We have talked about population health assessment in previous slides. We've talked about predictive models, and how to use a statistical analytic presentation to identify and stratify ad hoc reporting. It's something that can be used in developing reporting around sub-populations and subgroups of the population. Provider profiling reports are another output, as well as management dashboards that will go out to providers or to case managers themselves. And often have outcomes and evaluation that result from the statistical and analytic presentation. Now I have three examples of population assessments that I would like to review. There's an assessment of a Johns Hopkins health plan, a commercial population. There is an example of a population assessment in a self group of the population with diabetes. And then, there's an example of an assessment for the Johns Hopkins Community Health Partnership. So the first example is a population health assessment that was performed on a Johns Hopkins commercial health plan population. In this assessment, they wanted to characterize the demographic and morbidity patterns. And that helped them to guide their discussions about strategies that might improve the health of the population. It was used to eventually ameliorate the impact of patients with high disease burden on healthcare costs and utilization. So understanding the patterns of use, understanding the patterns of morbidity, help them to think about the disease burden and how they could then influence in a positive way healthcare costs and healthcare utilization an improved patient outcomes. So they have the following questions. How the populations compare to each other base on demographic and morbidity measures? Are there differences between groups within this commercial health plan population? Is it possible to identify the subgroups who are driving utilization? Or who may be at high risk for future utilization. What are the appropriate recommendations and strategies for improving the population's health, decreasing expenditures, and improving a patient's experience and satisfaction with the healthcare system? This table shows subgroups within the population. The far right-hand column contains data for all the members of the population, and the middle column shows data for the subgroup of the population being compared to the overall population. So in the entire population, 66% are female and the mean age is 37.5, a relatively young population. This is an employed population, 67.7% are employees, and the other percentage are dependents. The percentage with no medical claims, those who have never experienced any interaction with the healthcare system, is only 7.8%. And then the percentage of members who account for 70% of the healthcare expenditures, a relatively small segment of the total population is 17%. This table shows the mean adjusted clinical groups risk score, which indicates whether patients are healthier or sicker than the overall population in the two populations included in the table on the previous slide. A score of less than one means that the subgroup is healthier. A score of greater than one means that the subgroup is sicker. The table also shows the percentages of patients or members within this population who have no chronic conditions, one chronic condition, two chronic conditions, or three or more chronic conditions. These table shows prevalent rates per 1,000 in a population for selected conditions. So 10 in 1,000 employees have a condition or a diagnosis of anxiety in our overall population. In our sub-population only 8.8 per 1,000 have a diagnosis of anxiety. And from this table you can quickly get a good sense of the most prevalent conditions within the population. Hypertension, asthma, diabetes, and low back pain. And so then you can begin to focus strategies based on those most commonly occurring, or most prevalent selected chronic conditions. This table shows age and sex-adjusted prevalence ratios for selected conditions. So when an overall population, we are comparing a sub-population to the overall population. So for example, in anxiety disorders, for the age and sex after age and sex adjustment, to this prevalence rate, we have a group of people with anxiety disorders who are less at risk or less morbid than the overall population. And then in this situation, we're looking at depression in the overall population. The prevalence ratio is one. In this population, after age and sex, the prevalence is higher per depression than in the overall population. We also begin to look at the number of people within the overall population with conditions and what their mean per member per year cost are. So for people with a condition of asthma, the per person cost in the overall population is $16,063 per year. In our selected subgroup population their per member per year costs are less. The table shows that there are 198 people in the total population with diabetes and 563 with diagnosis of hypertension. Now let's look at chronic renal failure. Even though there is a small number of patients with this condition, the cost per patient per year is extremely high $222,000. So even though chronic renal failure has a small prevalence it may be a condition on which you'd want to focus because of it's cost and potential for poor outcomes Example 2 is in a health plan population. And we're looking across several different health insurance products at Johns Hopkins to develop a diabetes risk model. And in these scenario, the goal is to create a model that will be used to target and deliver clinical interventions to patients that are at risk for future hospitalizations. The model will help to identify the highest opportunity patients, who are the highest risk patients for population based interventions. The model is based on the patient, provider and health system factors that help determine health needs. And the model is grounded in clinical logic, meaning that there are clinical variables that are meaningful to members of the care team. When combined, they produce a risk score that is a robust predictor of cost and utilization. This model was based on the best available evidence. So in this example, they are devising and building a risk model for a diabetic population. Here, predictor variables were identified in their 2009 dataset. This chart identifies variables of interest. The first categorization is that of demographic variables, age, sex, zip code, region of the state, The John Hopkins health care line, a business or insurance product. In other words, was this a commercially insured patient with diabetes? Was it a Medicaid patient with diabetes, a Medicare patient with diabetes? One of our predictors was total cost of care. Pharmacy predictors, here they looked at the total count of medications. Whether the patient was using insulin, an insulin pen. Whether there was a lack of claims for certain supplies that would indicate the patient was engaging in self-management behaviors. Were there claims for glucagon? Was there untreated diabetes, PPI use. And you also have medication possession ratios, which we talked about earlier in the presentation. The adjusted clinical groups risk score, the overall risk score for future utilization is a measure of co-morbidity and how that co-morbidity impacts future utilization. Were there any enrollment lapses in the health plan? Any number of lapses? There could be three or more lapses in their 2006 to 2008, and 2009 datasets. Co-morbid conditions, besides diabetes, were there elevated laboratory values? Diabetic complications, hyperlipidemia, coronary artery disease, heart failure, renal, COPD, asthma, depression, substance use, or mental health conditions? In terms of utilization, what was the utilization of primary care visits, the number of visits, specialty care visits, the number and type of specialty care, hospital admissions, and ED visits? So these were all predictors that they developed from their 2009 data set. And they looked at hospital utilization in 2010. So how well did these predictors predict hospitalization in 2010? Again, their statistical strategy was regression. This is the distribution of their predicted probability scores across an entire population. There were 1,382 patients involved in their Diabetes Risk Prediction Model. And they chose some pilot primary care groups. The n was about 180. They broke down their population in terms of their overall risk level score. The estimated probability ranging from zero to one. And they found that they had a distribution that was skewed to the right. Most of their scores for most of their patients fell in the 0, 0.1 and 0.2 range. You can interpret those scores as a 10% or 20% chance, or probability, of admission or hospitalization in the future year. Those with a score of 0.3 to 0.4 were labeled as medium-risk. Those with a score of 0.4 and above were labeled as high-risk. So low-risk, medium-risk, and high-risk. They then, based on that model, developed a Johns Hopkins Community Health Partnership model. They were selecting patients for a program, but did not want to limit that program to diabetes. So they had as their goal, again, to create a model that would be used to select patients and to deliver clinical interventions to patients that were at risk for future hospitalization. Again, their model utilized the same four criteria as in the previous example. They wanted to identify the highest opportunity patients. They wanted to base their model on factors that they knew would help determine healthcare needs. It was clinically logical to the members of their team, and it was based on the best available evidence. In this particular program, they just wanted to identify the highest and lowest risk members. If you remember from our previous example, they had low-risk, moderate or medium-risk and high-risk patients. In this program, they wanted to identify 1000 patients who had Medicaid managed care insurance. And they distributed those patients by looking at the risk score. Again, a positively skewed population, where most of the population were falling in the low-risk range, 0, 0.1, 0.2 risk score. Those patients who were above 0.3 were considered high-risk. And they chose 1,000 as the cut off point. You can see the difference in terms of the number of people who were in the high-risk population. 1,000 chosen as high-risk out of a total population of 6,258. And 5,258 in their low-risk population. These people who have fell to the right of the cut point, or cut line would receive intervention. The average annual cost of the low-risk population was $5,400 and only 6% of them had an annual in-patient admission. As opposed to the high-risk population, those 1,000 people had an average annual cost of almost $30,000 compared to $5,000. And 47% of them had been admitted to a hospital at least once in the past year. Here's a complexity pyramid representing the patients in the previous slide. At the top are the 1000 high-risk patients, who had an average cost of $30,000 per person, per year. Of the moderate and low-risk patients, appearing here at the bottom of the pyramid, only 6% had an admission. The total cost of this group's care was about the same as for the high-risk patients. Though there were about five times more patients in the moderate and low-risk group. Their average cost per patient per year was $5000. So 76% of all hospital admissions were accounted for by people who fell in the top of this complexity pyramid. If we dig deeper, and look at the Johns Hopkins Community Health Partnership, J-CHiP, patient characteristics in the high-risk group, there were several medical and behavioral conditions that seemed to be very prevalent. Number one was heart disease. So 98% of the population had some indication. Through diagnostic claims, or electronic medical record data, or end organ conditions And/or modifiable risk factors. The rate of coronary artery disease, a condition that leads to heart attack, was 58% and the rate of heart failure was 32%. Of the modifiable cardiac risk factors, the rates of hypertension, smoking and high cholesterol were 84%, 71% and 52% respectively. Lung disease was also very prevalent in the population, 42% having asthma and 29% having emphysema, 28% had kidney disease. Substance use was prevalent, the smoking rate was 71%. 45% of the population had a substance use disorder and 29% abused alcohol. 49% of the population had Type II Diabetes. They looked at the geographic distribution of the high risk group. This is a map of the State of Maryland. Notice that even though they were focusing on six clinics within Baltimore City, there were several of their population of high risk patients, who live several counties away from Baltimore, at least an hours drive from Baltimore. So that was surprising to them that the distribution of patients was so widely spread. Looking at Baltimore City member locations they found that even though their clinics were designated within Baltimore City in a very confined area in East Baltimore, the patients were really spread throughout many, many zip codes throughout Baltimore City. So their High-Risk Group often traveled across the city to access Johns Hopkins' Medical Care. So we've looked at two examples of using risk-prediction models to select patients for population health intervention. Another way to select patients is through an acute-care utilization focus. This is done by identifying patients when a significant need occurs, such as a hospitalization or an emergency department visit. Or we could focus interventions on patients with a particular high risk condition, such as diabetes or chronic renal failure. Health risk assessment data often draw from multiple sources, including patient reports. So if patients report, for example, diabetes, poor self-management behaviors or poor lifestyle behaviors, they may be good candidates for a population health intervention. Additionally, we can have referrals by physicians or staff, and patients and employees can refer themselves to population health interventions. We will now review the objectives of Lecture a of Population Health Management Interventions. Prioritizing population health. We described population health data necessary for segmenting the population into risk-cohorts of low, medium, or high risks. We differentiated the key cohorts of a population by a degree of risk. We analyzed the root causes of risk in a population by utilizing various sources of data, including socioeconomic, behavioral, electronic medical record, and other demographic data. And we explained the processes and key decision points by which interventions are prioritized for segments of the population.
