Well, good afternoon, everyone, and welcome to today's webinar entitled Expanding the Boundaries of Healthcare Systems, Social Determinants of Health, and Community-Oriented Healthcare Delivery. I'm Dr. Angela Michelide, I'm the Vice President of Programs and Education at the American College of Preventive Medicine, and it's my pleasure today to welcome our speaker and our moderator. Dr. Elham Hattef will be presenting, and Dr. Antonio, also known as Tony Neary from the CDC, will serve as our moderator. Just a little bit about ACPM. We are a professional medical society of approximately 2,000 physicians dedicated to improving the health and quality of life of individuals, families, communities, and populations through both disease prevention and health promotion. Our members are licensed medical doctors who possess expertise in a broad range of healthcare skills, including biostatistics, epidemiology, management of healthcare organizations, research, and the practice of prevention in clinical settings. Our members work in a variety of settings, including hospitals, government at the local, state, and federal levels, the military, laboratories, businesses, and academia. I have just a few housekeeping pieces of information to share. At this point, the audience has been placed in listen-only mode, but we invite you to type your questions in the Q&A box throughout the presentation, and we'll answer those at the end. In addition, we'll be posting a reporting of this webinar on the acpm.org website. If you enjoyed it and want to share it with your friends or want to listen to it again, please do so. Of course, we want to know what you thought and how we can do better, improve our offerings in the future. So you'll receive a questionnaire, and please answer that because it helps us to develop our educational offerings. It gives me pleasure now to introduce today's moderator. Dr. Anthony Meary is a fellow of the American College of Preventive Medicine with more than 16 years' experience with the Centers for Disease Control and Prevention and the Indian Health Service. In these roles, he has led and managed large domestic, tribal, and international public health programs, many of which have been self-sustaining. He has a background in environmental health and chronic disease. Since 2015, he has directed the CDC's Preventive Medicine Residency Fellowship, managing a $1 million budget, three staff, and approximately 15 learners. Between July of 2020 and June of 2021, Dr. Meary's CDC telehealth efforts for the COVID-19 response by partnering to implement a strategic plan and a research agenda. He is currently on detail for the CDC Office of the Associate Director for Policy and Science to support the development of policy research. Dr. Meary, I turn the microphone over to you. Thanks, Dr. McIlyde, I very much appreciate it. And thanks so much for Stephanie Madrigal and yourself for putting this together and for ACPM for offering this opportunity to myself and Dr. Hetteff. My name is Antonio Nery, and I'm super excited to introduce Dr. Hetteff. Dr. Hetteff is a person I've worked with on and off for the last year or so on the response, and I am super grateful for the work that she does, and I'm very glad she's talking about such an important topic. Strategic data analysis and social determinants of health are really, really important and critical, especially in this day and age, and it's really, really nice that there's more focus on interventions related to social determinants of health, and in addition to identification of such issues. Dr. Hetteff is an assistant professor, she's a fellow in the American College of Preventive Medicine, an assistant professor in the Division of General Internal Medicine at the Department of Medicine, the Johns Hopkins School of Medicine. She also serves as a core faculty member for the Center for Population Health Information Technology in the Department of Health Policy and Management at the Johns Hopkins School, Bloomberg School of Public Health, and as the academic director of the Johns Hopkins General Preventive Medicine Residency Program. Dr. Hetteff earned her medical degree from Tehran University of Medical Sciences in Tehran, Iran, and her master's in public health from the Johns Hopkins School of Public Health. She completed the preliminary year in internal medicine at Yale-affiliated Griffin Hospital in Connecticut, and preventive medicine residency and chief residency at the Johns Hopkins School of Public Health. She then completed a clinical informatics practice pathway at Johns Hopkins. Dr. Hetteff is board certified in preventive medicine, public health, and clinical informatics. Her main field of interest is population health, social determinants of health, and health information technology. She studies the impact of social determinants of health on health-related outcomes using health information technology and big data. She's involved in several projects in this area. At the Center for Public Health, Population Health, and IT, she has served as the project leader on the development of population health frameworks and measurements for Maryland in collaboration with the Chesapeake Regional Information System for our patients. It was supported through the Maryland State Improvement Model and the Centers for Medicaid and Medicare Services. She also led the project to incorporate relevant available community population level data sources into the Veterans Health Administration's electronic health records. This project aimed to evaluate the health outcomes such as hospitalization rate at the primary care level while addressing social determinants of health. In addition, in collaboration with other faculty at CPHIT, she works on new methods of data mining and natural language processing to address social determinants of health using structured and non-structured electronic health records data and publicly available population level data such as the U.S. Census. With that, I pass the microphone over to Dr. Hetteff. Thank you so much for coming today and we look forward to your presentation. Thank you, Tony, for the nice introduction and thanks to ACPM for inviting me and everybody who has attended the call or is going to listen to the lecture in the recorded one. So I'm going to talk about basically using the data in the healthcare system and outside the healthcare system to assess the social needs and social determinants and how healthcare systems can change their focus from addressing clinical needs of the patients to a kind of more comprehensive approach to their patients' needs and to engage basically the communities that they serve in this process. Next slide. So I will basically talk about the data sources that are available, how we can assess the social needs and social determinants at patient, healthcare system and community level and how this approach kind of matches those famous three buckets of prevention that CDC introduced a while ago. Next slide. I think we all by now know the role that social needs and social determinants play in the healthcare system and the fact that they have a bigger contribution in the health of individuals and communities than any kind of medical care that we provide to our patients. These social needs and challenges are the reasons that may help disparities exist and unmet social needs and challenges basically cause different health outcomes for different subpopulations and they result in different mortalities among different subpopulations. Next slide. So the good news is that the development of EHRs and these cross-provider regional health information exchanges like the CRISP that we have in Maryland has provided the opportunity to develop some assessments of social needs and social determinants challenges and to design some interventions, whether it is at the point of care like assessing the needs that one individual patient has and referring them to social services or at the healthcare delivery system like identifying the clinics that are providing services for more socially challenged patients and hiring social workers to address those needs or at the community level, looking at the opportunities and challenges that exist in the community and developing some community initiatives to address the needs of the patients outside the borders of the healthcare system. Next. So a couple of years ago, some of my colleagues and I put together a paper and we basically talked about a framework, about like a conceptual model of how we can assess and address these social needs as providers and leaders in the healthcare sector. Next. This framework talked about the fact that assessing and addressing the social needs can happen at different levels. It can happen at the level of patient care. It can happen in the healthcare system, but a lot of the predictive risk modelings that happen in the healthcare systems. And I talk more about them or at the community level. We also talk about different stakeholders that are involved in this process from patients to providers, healthcare systems, payers, and even policy makers. And we touched on the data sources available, whether in the healthcare system, non-clinical data outside the health system, and also the data that comes from consumers. And I believe that this approach to addressing the social needs kind of mirrors those three buckets of prevention that CDC talked about a while ago. The first bucket of prevention was the type of interventions that happened in the clinical setting on a one-on-one basis. Addressing the social needs at the patient care level basically mirrors that first bucket. The second bucket was about the type of prevention that happened on a patient panel or maybe at a higher level in the healthcare system. Using predictive risk models to identify these populations very much mirrors the second bucket. And the third bucket was about the prevention happening at the higher level in the community. And assessing and addressing the social needs in the community mirrors that bucket. I like this framework because it shows that basically assessing and addressing the social needs is not that much different than the type of comprehensive approach that we have in preventive medicine to providing preventive care and also clinical services. So I will give you some examples of how we can use the available data at each of these levels to assess and address social needs. Next. So first I want to talk about data sources and what are the sources of data available, whether in the healthcare system or outside, and how we can use them to assess the social needs of the patients and to design some interventions to address those social needs and serious challenges. Next. So the first source of data, next, is basically data from electronic health records in the healthcare system, whether it is a structured data or the data that comes from the provider's free text notes. Next. A couple of years ago, we did an assessment at Johns Hopkins looking at the type of social and behavioral data that we can get from the EHR. Next. We wanted to get a sense of where we were with the type of social needs and social challenges that the providers documented in the EHR. To give you a little bit of background, Johns Hopkins started using APIC in 2013. Before that, different entities at Hopkins had these isolated electronic health records. And by 2012, Hopkins purchased APIC, and they started migrating all the data to APIC. And it happened at different stages. By 2016, all different entities across the institutions had access to APIC. So we basically used the data and the structured part of the EHR from 2003 to 2018, and we had about 5 million patients. And for the unstructured data, we focused on the last two years of 2016 to 2018. 2016 was when all the entities had access to the same EHR and they could document the information in the APIC free text notes, and 2018 was when we were doing the actual study. Next. Can you go to the next slide? So first, we put together a framework thinking about different types of social needs that we needed to focus on. At the time, Institute of Medicine and now Academy of Medicine released this report about the type of social needs that need to be assessed and documented in the EHR. Using that report, we developed this framework and we identified these domains and subdomains of social and behavioral determinants. The ones that I have circled here are the ones that we focused on to assess the data availability at Hopkins. Next. This very busy table basically shows where we were at the time with documenting the social needs of the patients. You can see that for some of the variables, like the location of the patients, many patients had the information. And this is important because if you know where your patient lives, then you can link that data to the population level data. And there are a number of different ways of using this linked data to assess the social needs and social challenges of the patients. And I give you some examples. For some other variables, like race, we were doing better. But for ethnicity, like half of the patients had information on ethnicity. Then when we got to some behavioral ones, like alcohol, less than 10 percent of the patients had the information. And also for something like smoking status, which is a kind of common part of the social history, there were only 30 percent of patients having information on their smoking status. And at the time, we looked at other health systems and other reports and other health systems weren't necessarily doing that much different from Hopkins. Next. So then we used the Z-code or ICD codes that are available to providers and like this ICD codes for clinical conditions, there are the Z-codes available for documenting the social needs. And we wanted to see if the providers really used those Z-codes. And for social connection, housing and income and financial resource strain, we looked at the number of patients that had any of those Z-codes. And you can see that almost all of them had less than one percent of the patients had any of these Z-codes. It doesn't mean that the Hopkins patients didn't have any social needs, but just using these ICD codes wasn't really a common practice at Hopkins to document their social needs. Next. So then we did an update version of this study last year and we selected again a number of these social needs and did a very comprehensive assessment of documentation of those social needs using the ICD codes in the system. We wanted to see if the practice of the providers changed in a couple of years in terms of using more social needs. Next. This table basically shows the characteristics of the general population and also the population that had any of those social needs documented for them in the EHR. We focus on these four categories of social needs. And if you look at their, for instance, their age distributions, younger patients had more social needs comparing to the older ones. As expected, African-Americans and other minorities like Asian-Americans had more social needs. Also, the health care utilization was higher among the ones that had social needs, like any outpatient clinic, ED visit or inpatient was higher. And you can compare these numbers to the first column that is basically the health care utilization for the total population. And also patients came from the different counties across Maryland. And for instance, the patients that came from Baltimore City, 2.5 percent of them had some social needs. I want to mention that because the practice of assessing and also documenting the social needs is not something common currently in the health care system. These numbers really don't show the prevalence of these problems. They basically show the patterns of assessing and documenting the social needs. We wanted to see which type of patients were more prone to be asked about the social needs and which type of patients had more information related to the social needs for them in the EHR. And this table basically shows that. Next. We also, as I mentioned, looked at the patterns of health care utilization for these patients. We looked at the number of outpatient and ED and inpatient visits that the total population had, the first row, and compared it to the utilization patterns for the patients that had any kind of social needs. And you can see that, for instance, outpatient visits for patients with social needs was around, on average, eight visits. But for the total population, it was around three. The same thing, emergency department and inpatient was really high for the ones that had any social needs. We also looked at the most common ICD codes that were used. And these three ones are the ones that were the most documented ICD codes. And you, again, can see that the ones that had these ICD codes, the ones that I circled, compared to the general population of Hopkins had higher emergency department and inpatient visits. Next. So because information in the structure EHR is not that much, the next question was that, are the providers documenting these social needs in the unstructured EHR, the free text notes? Next. So we started digging into the free text notes. And to do that, we basically selected a number of social needs and social connection isolation, housing issues and income. Next. And we started developing these patterns, these different phrases that basically described those social needs. And because we don't have any terminology or any kind of standard ways of describing these social needs, it was very challenging to develop those linguistic patterns. So we, again, listed all the available ICD codes, looked into the public health surveys and instruments, and also looked into the literature and did some manual annotation of the notes to come up with a number of different phrases that would describe specific social needs like housing issues. Next. Then the way that the NLP works is that you use those phrases that you develop and the software engineers develop these algorithms based on those phrases and those algorithms review all the free text notes that were available in our system. At the time, there were about 20 million notes and they identified the notes that had any mentions of social needs, those selected social needs. And here I'm just showing the number of patients and the percentage of patients that had at least one free text note with one mention of one of those social needs. And for social isolation, it was about 2.5 percent of patients, housing issues, 3 percent of patients and less than 1 percent of patients had at least one note with mentions of social needs in there. And you can see that the numbers are not that high. But if you compare them with the first column, which is basically the data that we got from structured EHR, you see that the unstructured EHR or free text notes had many more mentions of these social needs. It means that providers, if they assess the social needs, they would more document them in the free text notes rather than the structured EHR. We also looked at the type of providers that documented social needs. Physicians were the ones that documented them the most and also the type of notes that had any mentions of social needs and partners notes were the ones that had the highest number of mentions of those specific social needs. In the next phase of the project, we basically repeated the same approach, but we basically tested the algorithm that we developed at Hopkins in other health care systems. We collaborated with Kaiser Permanente, both in the Mid-Atlantic region and also Southern California. And each of these three health systems selected their own patient population based on the criteria that we identified, and then they developed their gold standard and then they ran basically those NLPA algorithms in their data. We wanted to see how these NLPA algorithms were functioning in different health care systems. Next. And in this table, you can see that the last two rows of the table show the sensitivity and specificity of the data. It basically shows that the NLPA algorithm that we developed at Hopkins was doing relatively well at Hopkins, but also if you go to the second and third column, look at the performance in the Kaiser Permanente Mid-Atlantic and Southern California, the performance of our algorithm was relatively well, especially in Southern California. So this was very important for us because for the first time, we basically developed an algorithm that could go beyond one health care system and one type of EHR and kind of get similar information from other types of EHR. So it shows that these algorithms can be generalizable and can be used in different health care systems to identify the social needs of the patients in the free-tax EHR. Next. So I guess this result shows you that still documenting the social needs of the patients is not really a common practice in the EHR. That might be the reason that a lot of health systems and also in the industry, many people are using their non-clinical data outside of the health care system to assess the social determinants or the characteristics of the neighborhoods that the patients come from. And they kind of use that as a proxy to assess what type of social needs the patients might have. One of the famous ones is, as you might have heard, this American Community Survey that Census runs every year, and they basically survey about three percent of the population. And they release a lot of information on different social determinants, called the characteristics of the neighborhoods, at different geographic levels from state to county, census tract and census block group. Next. Next. And then there are also some other data sources, such as the data that comes from consumers and even the data from social media that might have some information about the social needs of the patients. But there are a lot of challenges and a lot of discussions about the quality and validity of that data. I want to move quickly to give you some examples of with this type of data that we have, how we can assess the social needs and design intervention to address those needs at the patient level, as I mentioned at the beginning of the presentation, at the health care system level and also in the community. At the patient level, this diagram is actually from an R01 that we recently was awarded. And the idea is that we want to use the patient and population level data that we have to develop a social risk score and a decision support tool in the EHR. And that decision support tool would help the providers to identify the patients that are in need of the social, that have social needs and challenges. And then we would develop a closed loop referral in collaboration with Maryland Health Information Exchange and connect those patients to community organizations. And then the providers could refer patients to community organizations. We built a test that in selected clinics in a clinical trial and then assess how this whole system has helped to identify patients in need and how addressing those social needs impacted their health outcomes or patterns of health care utilization. So this is an example of basically making use of patient and population level data to identify the social needs and address them at the patient level as the first bucket of prevention that I talked about. Next. But there are also ways of making use of this data at the health care system level, and one of the famous ways of doing it is basically developing these risk predictive models. These models have been used in the health care system in the US for a very long time. And for the most part, they were relying on claims data and some clinical data from EHR more recently to basically identify patients that are at higher risk for having poor health outcomes. And these predictive models really help the health system to target the specific subpopulations of patients to and to develop some interventions. But there's not that much use of social and behavioral data in these predictive risk models. Next. And actually, there are a lot of studies looking into how we can make use of this social data and predictive model and the findings kind of mix the ones that come from academia are not that optimistic, especially about using population level data in the industry. There are many companies that are using population level data in their predictive risk models, and they claim that they are able to identify patients at higher risk by just using or including social and behavioral data in their predictive risk models. We did a pretty comprehensive review paper on challenges and opportunities of using this social behavior data in predictive models. Next. And we basically talked about what they bring to the table and what are the pitfalls or challenges of using them and how we can basically improve the use of this social behavior data in predictive risk models. Next. We also made a number of recommendations about how to address a lot of challenges related to the use of this data in the predictive risk model. Challenges related to the patient privacy and ethical issues of which type of social data you want to use, how far you want to go with including this social data in your risk predictive models. There are a lot of technical challenges as well, especially because the data comes from different sources. Some of them are at the patient level, some of them are at population level. Even using data in predictive risk modeling from so many data sources has a lot of challenges, and there are a lot of people working on which specific modeling techniques you can use to basically include different types of data. And also challenges related to the fact that there aren't any best practices or guidelines about how to include this data and how to deal and handle the data in predictive risk model. So if you're interested in this topic, we have made a number of recommendations about addressing each of those challenges. Next. So I want to give you a couple of examples about using patient and population level data for predictive risk model. The first one is the work that we did with VA. We had access to their EHR data from the patients that were in their primary care network, and we linked that data to population level data from census. And we basically developed this neighborhood socioeconomic status measure that takes into account a number of the variables from the community. And next slide. We wanted to see how the combination of these patients and population factors would impact the patterns of hospitalization for these patients. Next. First, I want to quickly show you in this paper that the hospitalization rate was a little bit lower in VA comparing to the general population. Next. But when we did modeling and we looked at how characteristics of the neighborhood or this NSES, which is the neighborhood socioeconomic index, impact the hospitalization, we basically compared the neighborhoods that were in the second, third, and fourth quarter and had higher SES scores to the ones that were in the first quarter that had the lowest SES score. And as expected, you see when the socioeconomic status of the neighborhood went up, basically the hospitalization rate went down. Next. We also did a similar study at Hopkins using both patient and population level data from the EHR and also from census. Next. And this time we developed this area deprivation index, which is another composite measure of assessing the characteristics of the neighborhood. And we had about 1.2 million patients in the EHR. Next. And we looked at how the characteristics of the individual patients and also where the patients impacted their patterns of healthcare utilization. We did the modeling both for the overall population and also for the Medicaid population. And you can just look at the rule that talks about the neighborhood characteristics, the ADI. We compared the ones that were at higher ADI, which means that they were at more disadvantaged neighborhoods, to the ones that were living in the 10th percentile or lower ADI, which was the ones that were in the less disadvantaged neighborhood. And you see when the challenges in the neighborhood goes up, both for the general population and also for the Medicaid population, they had a higher odds ratio of utilizing the healthcare services. The same thing, we compared the Medicaid and Medicaid patients to the ones that were in the commercial insurance. And you can see that the ones, especially the Medicaid ones, had higher utilization markers comparing to the commercially insured patients. So, this is, again, another example of making use of patient and population level data, developing this predictive risk model to predict patterns of healthcare utilization for a specific subpopulation of patients. Next. We also did the same thing, again, using claims data at Hopkins and linked it, again, to census data and make a community survey. We specifically focused on priority partner patients who are the Medicaid population at Hopkins. Next. And this time, specifically, we looked at patterns of ED utilization. And again, you can see that for the Black people, comparing to the whites, they had higher utilization. And again, if you compare the patients, this time, their reference was neighborhoods that had the highest ADI, which means that had the most challenges. Comparing to the most challenged neighborhoods, the ones that had higher SES or had less challenges, they had lower ED utilization. Next. We also mapped this ED utilization. Next. The right map basically shows patterns of utilization around Hopkins for different neighborhoods in Baltimore. And the left map shows the ADI distribution. And in the left map, you basically can see that famous butterfly effect that East and West Baltimore had higher ADI, which means had higher socioeconomic challenges. And if you compare it to the left map, you can see that most of the ED utilization at Hopkins came from a handful of neighborhoods in East Baltimore that were socioeconomically challenged. This type of data helps the healthcare systems to basically identify address populations and to be able to design some community-level interventions that address the needs of those patients. Next. So, that brings me to the third and last bucket of prevention or last way of using this data for assessing the community challenges and developing interventions that address the needs of the patients in the community. This is a study that we again did with VA. This time, we focused on the patients that get in services at the VA primary care headquarter in King County, Washington in Seattle area. Next. We wanted to see how the housing challenges that the neighborhoods had where the veterans live impacted the pattern of hospitalizations for those patients. Next. We again did the same thing that basically chose a number of housing variables from census, and we linked them to the patient-level data from the EHR at VA. Next. And we mapped these housing challenges, and each of these maps, the first one on the left shows basically housing instability, then vacancy rate, houses with no plumbing, and houses with no fuel waste heating. And you can see that the ones that are circled are the area around Seattle, and those were the neighborhoods that had the most challenges. Next. And then we mapped hospitalization again, and you can see that hospitalization, if you compare this map to the other maps, you see that most of the hospitalization basically happened in the neighborhood that had those housing challenges. Again, VA could use this mapping to identify the neighborhoods that had the highest challenges, and those would be the neighborhoods that they would focus on to basically direct their population-level interventions. Next. I want to go quickly through a couple of examples of how you can use this social data on a population and patient level to address challenges related to the management of COVID. Next. We did a study at the beginning of COVID, and we basically used the number of COVID, reported COVID cases in different states, and we linked it to the data on social media. Next. And we also used the data and we linked it to the data on socioeconomic challenges of the neighborhood from census, and we wanted to assess whether there was any correlation between the number of COVID cases with social and economic challenges of the neighborhoods. Next. And we did this at the beginning of the study, at the beginning of the pandemic, that many people were talking about COVID being higher among the neighborhoods with more challenges, but there weren't any quantitative data available on that. This table basically shows the correlation between living in a neighborhood that has socioeconomic challenges with the number of COVID cases. So, we looked at the states that had a number of COVID cases at a zip code level, and you see at the beginning of the pandemic, early May, that not many states had a lot of patients tested. You can see that in Illinois, Maryland, and South Carolina, living in a more socioeconomically challenged neighborhood meant that the patients would have, there would be a higher number of COVID patients. And when we adjust the correlation, the numbers were the same and they were significant. These are the states that at the time had a higher number of patients being tested. The rest of the states that really didn't show that relationship didn't have many patients tested. By the end of May, when the number of tested patients increased, you can see in all states, higher socioeconomically challenged neighborhoods had more COVID cases. So, they were all kind of proving the hypothesis that we had, except for Virginia, that the relationship was reversed. When we looked at the patterns of COVID testing in Virginia, it was interesting that in the neighborhoods that were more affluent, they had more patients tested. So, this reverse correlation in Virginia was because of the fact that the more challenged neighborhood didn't have enough patients tested, so they didn't really identify their patients. Next. So, this was basically a very quick overview of the type of data available in the healthcare system and outside healthcare system to assess and address the social needs and some of the challenges that we will face when we are using this available data. Although there are a lot of opportunities, I want to also acknowledge the fact that the data brings a lot of challenges. There are challenges related to developing best practices and standards or tools that can make use of this data. Also, there are even challenges related to assessing and documenting these social needs in their health system. And there is a really cost involved in that for healthcare system, and there is a lot of labor burden on providers and health systems. And also, challenges related to the data itself, that the data is not interoperable yet. There are challenges related to the validity of the data or even the confidentiality of the data that is getting collected that needs to be addressed. Next. But although these challenges exist, I think we all acknowledge that there are growing pressures from payers and policy makers on providers and healthcare systems to offer better services at lower cost. And if the clinicians and health systems and providers really want to meet those quality measures, they really cannot not to take into account the social and behavioral needs of their patients. There is also the need for developing evidence of best practices in this domain, but nobody can anymore develop these interventions and these analytic platforms without taking into account the existence and the need for social and behavioral data in this approach that we have to patient care and population health. And with that, I would like to thank my colleagues. Next slide. And I would like to thank everybody for listening and I'm available to answer the questions. Thank you. Thank you, Dr. Hintaf for your presentation. As a reminder, you can enter your questions in the question box and we will work on filtering them and screening them and getting over to Dr. Hintaf in the interim. Dr. Hintaf, I had a couple of questions for you. Sure. I was wondering if you could talk a little bit about the modeling that you're doing and then how that translates into policy levers or intervention level factors, and particularly at the provider level. I think that's one of the difficult parts that I face in practice and being able to not only see that someone comes from a high-risk area or a difficult situation, but how to address them. I was wondering if you could talk about the modeling and how those two mix into a practice. Sure. I think in terms of the practice at a patient level, the way that these models have been developed traditionally, they've been really developed for the leadership level or the management level. One of the examples of using it at the patient level is that example that I gave for that social risk score. What we are developing with that social risk score is basically a model that predicts or identifies the patients that are at higher risk of having the social needs or having poor health outcomes. Then that model generates a risk score that is available for each specific patient. The provider basically can use that risk score to decide whether this patient is at higher risk of having social needs or lower. I think my point is that with health informatics tools that are available, there is the possibility of bringing that model that is developed at the population level to develop very specific indices for specific individuals in the system, specific patients in the system. Then the providers can use that specific score and their clinical judgment to make a decision whether they need to do any more assessment of the needs of the patients or whether they need to do some referrals. I want to acknowledge that for the providers to do that, there needs to be a whole system around that. The clinics need to know who is supposed to do the screening or when they do the screening and identify the patients, what are they supposed to do with them? There should be also a referral system to link those providers to community-based organizations. As long as these links are not developed, asking a provider, a standalone provider to just do the screening and to figure out what to do with the patient when they identify a social need, I don't think that's realistic. That might be the reason that a lot of these tools that have been developed and implemented in the EHR haven't gone that far in terms of becoming like a common practice in the clinical setting because there are a lot of missing links after identifying the patient or screening them. And then further, I wonder if you could talk a little bit about the difficulty in being able to obtain that information, particularly with electronic health records that may not be synchronized or at odds with each other. Yeah, I think one of the major challenges is exactly the fact that the EHRs are really silos and a lot of these social needs, even when you identify them, they are not the type of needs that the healthcare system is really trained to address them. At the end of the day, you find somebody with housing needs and you really need to refer them to social services. One of the challenges is that even when there are some ways of communications between the healthcare system and those entities outside the health system or in the public health sector, there are a lot of miss... The EHR systems basically don't help those communications. There is like this cut or there is a missing link in this whole communication because the information from the EHR doesn't get out of the system. One of the ways of addressing it is what, for instance, CRISP is doing in Maryland, that they are trying to make use of this health information exchange in Maryland to connect those community-based organizations to each other, or even connect them to the health system. Then the EHR has some output that goes outside the health system, or even sending some reports to the public health agencies. The EHRs are trying to develop some sort of screening tools and management tools that are relatively similar. They might not be exactly the same between, for instance, Epic and Cerner, but they have the same structure. Health information exchanges can make it, for instance, mandatory for the health systems to report those instances of social needs in a common format and basically communicate them with the public health sector, something that Maryland is doing for a lot of their clinical indices. Hospitals and healthcare systems by law in Maryland are required to report some of their clinical information and send them in a structured standard format to Maryland Health Information Exchange. Reporting of social needs can be something similar to that. Certainly, Maryland is one of the early adopters of the Accountable Care Organization model with competition. I was wondering if you could talk a little bit about that in comparison to your other locations, particularly with the Affordable Care Act and ACOs, with this idea of addressing population-level health in the Medicare Shared Savings Program. I know it extrapolates it further, and you can say no, but it's a question I have, particularly because of your institution and the population you're working with. You can always defer. Sure. I think one of the experiences that Hopkins has had has been this contribution in this accountable health community more than that CMMI has had, that some of the selected primary care entities in Baltimore and also Baltimore Health Department have been required to participate in this pilot model of identifying the social needs of their patients in addition to their clinical needs and referring them to their community-based organizations. The idea is that basically these patients would have better health outcomes and they would have lower costs. For health systems that, like in the Maryland model or in this total global budget model, that the health systems get basically paid a bulk of money to take care of a patient population, having this approach of addressing non-clinical needs is a way of managing those costs and keeping them at the lower end. So, it sounds like they seem to be ready adopters, right? A more willing and interested population of adopters. Does that sound about right? In terms of this high-level modeling and being able to identify those, it would seem that there'd be a ready uptake of that. I think the major factor in being ready to adopt it is the fact that the payment model very much supports a more population health approach to providing care. If your health system is getting penalized for having super-utilizers or having higher readmissions, or if you basically have a fixed budget to deal with, no matter how many encounters the patient has with the health system, then that provides a lot of incentives for the health systems and providers to look into other ways of providing care and to have more population health approach. And if the health systems want to keep their costs under control, one of the major areas to make the improvements is basically looking at non-clinical factors that contributed into this super-utilization of services and to go beyond just patients that come to their clinics, try to identify patients that are in the community that might end up being super-utilized. So I think one of the major driving factors is just basically the payment models and the way that the health systems get reimbursed. Thank you. That's a complex, it's a very complex situation and I very much appreciate your work in that area. I wonder if we could switch a topic question a little bit to talk about large data analysis and certainly with the bigger data sets you have, particularly with those participants who might not be able to maybe even understand but look at the different areas of storage and computational capacity and analytics for these large electronic health care systems as they grow. I mean, they grow every day with number of users and volume of use. I was wondering if you could talk a little bit about your work in that area as it relates to the topics you've talked about. I think some of the work that we have done, yes, they really need some health systems that have that computational capacities, like the work that we do with natural language processing is still, it needs very big servers and a lot of computational power in a health system, which might not be necessarily available for a smaller health system or a number of providers working in a smaller setting. But I think beyond like these development phases, there are a lot of advances happening in making these health IT tools something more accessible and manageable for the providers. And I want to say that there are a bit like not very advanced or very fancy EHRs, EHR tool, but something kind of more modest in terms of the capacity. There is still a lot you can do with the data that's getting collected in the EHR. I think it's more important to have a more population health approach to that data and to have a sense of what one can get from the data and how they can use it in practice, whether at a patient or population level. Some of the more complicated developments, yeah, they require to have a larger health system. And I think there are even companies like what Google or Amazon is doing, I think they are making it more manageable or doable for a smaller healthcare system to use these health IT tools with maybe a lower capacity in terms of the technical tools that they have available. And particularly, if you could, you can always, you know, expand a little bit on natural language processing for the audience, because I think that that's an interesting approach in which to automate some of these processes for data identification. And I'm not sure analysis is the right word, but I wonder if you could talk a little bit more about that. I know we're coming close to the top of the hour. Sure. So with natural language processing, the idea is that you find some systematic ways of computer algorithms to go to a very large number of clinical nodes to identify very specific information. You can start with something very simple, like have very specific phrases that they develop some rules around them, software engineers, and the computer just uses those rules. And if anybody has very specific mention of a phrase, they identify that. There are kind of more sophisticated ways of expanding on those rules. And for instance, including negations in a model that I don't want you to just find somebody with housing instability, I want you to find somebody if the sentence says that they don't have housing instability, or I want you to find somebody that their family had housing instability, but not the patient itself. So there are different ways of kind of expanding on those rules and developing these algorithms that are more sophisticated. And that is like the natural language processing part. There are some machine learning techniques that are more complicated, and they are basically these statistical models of identifying somebody that has a higher probability of having a specific social need based on the specific phrases that you have included in that model. So those machine learning techniques are more complicated. Some of them are more like a black box. Some of them are more opaque, and some of them are more clear about what they do. But the main concept is that you want to have a number of specific phrases and develop different, maybe variation of those phrases and develop different rules that the computer can, closer to human brain, review these large sets of data and more accurately develop and identify the patients that have those specific social, those specific needs for incidences documented in their pretext maps. Thank you so much. It's such a complicated process, but I think it's really germane towards a large data set analysis, and particularly for those low resource or resource-lacking settings for analytics and data abstraction and being able to find some, look at large data sets, whether they're census or other EHRs as well, but be able to use a backend models or approaches in which you can make that data more palatable and interpreted. Thank you so much. Well, thank you, Dr. Hedeff. I very much appreciate your time and effort today. With that, I think we're going to close our presentation and turn it over to Dr. Mikolaj to help close us out. Well, thank you both very much for a fascinating session. We had a lot of participants very interested in this topic. We'd like to alert you on about a few upcoming webinars, Innovative Technology in Preventive Medicine with Dr. Colby, will be held on September 21st at 2 o'clock Eastern, and another webinar on colorectal cancer advocacy from an organization called Fight CRC that is working to get both in-home and colonoscopies paid for by health plans or by federal programs. That'll be on September 29th at 1 o'clock. Please be on the lookout for the event survey and for more information about more than 250 offerings to help build your CME portfolio, please log on to the ACPM website under education and events continuing medical education. And with that, we'll wish you a very good day and hope to see you soon. Thank you.

Expanding the Boundaries of Healthcare Systems

Social Determinants of Health

Community-Oriented Healthcare Delivery

Data in Healthcare Systems

Electronic Health Records (EHRs)

Assessing Social Needs

Predictive Risk Models

Privacy and Confidentiality Issues

Collaboration for Health Outcomes