Leveraging Data-Driven Advanced Analytics and Artificial Intelligence Technologies
DR. CHRISTINE HUNTER: Welcome to today’s OBSSR Director’s Webinar, titled, “Leveraging Data-Driven Advanced Analytics and Artificial Intelligence Technologies to Address Social and Behavioral Determinants for Health Equity.” I’m Christine Hunter, the Acting Director of the Office of Behavioral and Social Sciences Research at the National Institutes of Health. I apologize that I’m not on camera; my web camera has decided to stop working about 30 minutes ago. So, before I introduce today’s speaker, a few housekeeping items I want to cover. So first: Today’s webinar is being recorded, and the recording will be available in about 1 month on the OBSSR website at obssr.od.nih.gov.
Today’s presentation will be followed by a question-and-answer session. Throughout the talk, all attendees are muted, but the chat feature…and the chat feature is disabled. So, questions and comments will be taken via the Zoom Q&A feature. To ask a question or send a comment, click on Q&A at the bottom of your Zoom screen, type in your question, and send. You have the option to “like” other questions to avoid duplicate posts. The most-liked questions will move to the top. Feel free to send a question at any time during the webinar. Following the presentation,
OBSSR’s Dr. Beth Jaworski will facilitate the Q&A session and ask your questions to the presenter. So, with that, I’m pleased to introduce today’s presenter, Dr. Irene Dankwa-Mullan. Dr. Dankwa-Mullan is a nationally recognized industry physician and scientist, a health equity thought leader, scholar, and author, with more than 20 years of diverse local, regional, national, and global leadership experience in health care systems, businesses, and the community.
She is currently the Chief Health Equity Officer and the Deputy Chief Health Officer at Merative, formerly IBM Watson Health. Her current research strives to develop and evaluate data sets, real-world data, algorithms, and…and algorithms as inclusive technology—so, artificial intelligence and machine learning–driven technologies—to empower health providers, patients, and their families. A priority is advancing technologies to promote social good and equity. She supports inclusive and participatory engagement with communities
and stakeholders, and she also helps teams with modeling complex decisions associated with health equity and social determinants of health. Dr. Dankwa-Mullan has engaged in the implementation and evaluation of data and evidence studies, including social, legal, and ethical implications of use of these emerging technologies. She was formerly Deputy Director, Extramural Scientific Programs, in the National…at the National Institute on Minority Health and Health Disparities and played a key role in promoting strategic trans-NIH and Federal efforts. Dr. Dankwa-Mullan has published widely on health disparities, evaluation of artificial intelligence and machine learning technologies, including the integration of health equity, ethical AI, and social justice…principles into AI/ML development lifecycle.
It is my pleasure to welcome Dr. Dankwa-Mullan to present today on leveraging data-driven advanced analytics and artificial intelligence technologies to address social and behavioral determinants for health equity. And with that, Dr. Dankwa-Mullan, I will turn it over to you. Thank you. DR. IRENE DANKWA-MULLAN: Thank you so much. Thank you, Dr. Hunter, for the introduction and for the kind…invitation to present at this webinar to share my insights on this topic that I am so passionate about: how we’re leveraging data and technology. So, I’m really delighted to be here. I’m grateful for this opportunity to provide some perspective in this space, as well as some of the current and emerging research. So, this is…I want to emphasize that our new company, Merative,
is an extension of our ongoing efforts, and so it’s an exciting time for all of us, you know, to…to build on these technologies. And so next slide is the outline of my presentation today. I’m going to provide an overview of the role of data—so, big data, advanced analytics, AI/machine learning—in the domain of social and behavioral health. So, I’ll talk about our efforts
to advance health equity, racial justice, building inclusive technologies. I’m going to provide an overview of AI and machine learning and some of those terms and concepts that we use in this space to provide a foundation for my presentation and also help with that discussion. I will highlight some recent innovative efforts to demonstrate how we can integrate behavioral social determinants data. And…and briefly discuss a general framework around addressing bias, which is a huge deal and effort that we…we’re always thinking about bias in AI algorithms. And I’ll conclude with some thoughts for the future, including community and stakeholder engagement.
So, before I talk about this, I actually joined IBM Watson Health, now Merative, about 6 years ago as part of a clinical and health care experts team to lead and promote clinical evidence and evaluation. So, in my role as Deputy Chief Health Officer, I provide the subject-matter expertise, clinical expertise for scientific evidence to…to prove effectiveness and value technology and solutions. And in my role as Chief Health Equity Officer, I…I also provide strategic leadership support, subject-matter expertise for how we can think about health equity to ensure inclusive technologies and data diversity. And so I…I think often about how we design better solutions
in health IT for health equity. I think about the patients and populations, the entire health ecosystem, how we can optimize or leverage the data…the data assets that leaders and communities, researchers have entrusted in us in an ethical manner, with…with privacy, with transparency, and build, you know, trustworthiness. So, some of these examples...how we’re working…we’re working with our partners, we’re working with collaborators, really to design robust data representations. And…and my goal is to make sure that we’re capturing the complete life experiences and understand these data points and how they impact health outcomes. So, I mentioned that this is an extension of our…of our efforts, and some of the programs or initiatives that we have implemented just in the health equity space is, for example, building inclusive technologies and promoting inclusive language. And this program really included identifying discriminatory terms in…in our technology. So, terms such as, “master” and
“blacklist” that were used in that industry have been removed from use, replaced by inclusive words. We think about bias in a machine learning algorithms and have developed cold audits on how to mitigate those bias. We published a framework called TechQuity, which is promoted across the continuum of the technology design and development lifecycle. And this concept really calls for technology and businesses to be accountable for the active promotion of health equity. We’re working with partners to build health equity dashboards and…by enhancing their own data or existing health solutions, we’re…we think about integrating socio-demographic factors in health equity and metrics. And so, really promoting…technologies that may include AI
or machine learning as a strategic lever to make health care more efficient and more equitable. One of the things that we…I was also excited about is design justice to promote racial equity in design, thinking about representation, which really matters. So, a whole host of really great, exciting work around really leveraging technologies and promoting health equity. And then I do want to touch on AI ethics because there are social, economic, legal, and ethical implications for leveraging data and machine learning solutions in AI. And so we really take AI ethics seriously, and AI ethics is really a multidisciplinary field of study, and the goal is to understand and optimize AI and machine learning’s beneficial impact while reducing risk and adverse outcomes for all stakeholders in a way that really prioritizes our well-being, our human agency. And examples of AI ethics are illustrated in this slide, where it’s data responsibility,
it’s explainability, it’s…you know, robustness in our data, transparency, moral agency, aligning your value with the communities. And so, we…and I’m pleased to say that we led the development of a framework for integrating health equity racial justice principles into the development lifecycle with experts and partners in the field. And so, it’s a framework that can be used by researchers and developers and stakeholders to really assess the impact of their AI tools and technologies to ensure equity and racial equity, social justice, is prioritized. So, on the next slide, I just want to…a little bit about Merative. It’s really an extension of
what we’ve been doing at IBM Watson Health. We do support the health care industry and clients who deliver health and human services, working with them, leveraging technologies and their data to improve health. Not only in cost and in quality but, most importantly, in innovation and outcomes. And our mission is the same, as I mentioned, but we realized that our need is greater. Our need is so much greater, given the ongoing efforts of the pandemic and the global recession and overall economic hardships that have really been seen, you know, across communities. And
our community responsibility also reflects our expansive footprint in the health care space, and so it includes our social responsibility to our employees and our workforce and our clients. And so, how can we collaborate for innovation to address these really critical societal needs in the communities in which we operate with a culture of ethics and integrity and promoting trust and really placing people at the center of...of health? And so, this is…slide shows some of our product family. We do have Health Insights—so, these are end-to-end analytics and data solution that’s designed to manage population health and health care program performance. We have Social Program Management, and that helps with health and social program administration at the point of care, including benefits management, family support programs, child welfare, and…and all those programs that I used within the health and government social program sector. Micromedex is a clinical decision support tool that integrates evidence-based drug and disease content. We have our MarketScan—some of you may have heard
about it. It's this integrated, patient-level data reflecting real-world continuum and cost of health care, and it’s one of the largest proprietary collection of de-identified U.S. patient data available for health care—over 270 million lives in there. And then Clinical Development in…in helping with clinical trials. And Merge is actually our Enterprise imaging and AI-enabled solutions for radiology, cardiology, and to manage imaging data from a centralized platform.
So, data is really, really transforming every aspect of our world. It’s…every aspect of our lives. We use data to make decisions in so many ways. And in the health care space, data is always being leveraged to allocate resources, to target interventions, to identify populations at risk, and so much more. And data points are used in just about every domain of health care, including especially, behavioral and mental health care for—example, in clinical decision making, in understanding treatment pathways for optimal outcomes, for self-care and management of comorbidities, and more. And so, you know, the question is: Are we using the right data? And I want to start, on my next slide, to really level set with some definitions for AI and machine learning technologies. So, what do we really mean when we talk about AI, machine learning, deep learning, natural language processing—all of which are being used? How are they being used, and how we integrating the data, including those relevant behavioral and social determinants of health to surface those meaningful insights for improved health care.
And so, this slide really talks about…so, as you can see, the…an AI system is a system that can make predictions and recommendations or decisions that influences our physical or virtual environment. And so, they’re typically trained with huge quantities of structured or unstructured data. And they may be designed to operate with varying levels of autonomy—or none—to achieve your defined objectives. So, as you can see, AI basically is composed of machine learning and...or natural language processing. So, AI…machine learning is a subdomain of AI,
and it refers to that family of algorithms that can identify patterns in data. And so, machine learning can make predictions based on patterns in huge amounts of data. Natural language processing is basically another system of algorithms that can understand language and syntax and interprets the written language. Deep learning is a subdomain of machine learning, and natural language understanding is also a subset of natural language processing that deals with machine comprehension, and it…it’s defined by how machines understand human language and behavior. A lot of this is also being used especially in the mental health space. But my talk today I’m going to focus on the AI and machine learning and deep learning domain. But just to give you a sense of the…or…you know, what AI is and the various terms
that are being used. Next slide. And so, really apologize for a lot of text in this graph, but I want this figure to help illustrate the spectrum of advanced analytic applications and technologies that are used to generate insights from big data. And so, the methods can range from less complex but advanced analytics, such as descriptive analytics, that do not involve artificial intelligence to more advanced, more complex analytic methods that will…does involve deep machine learning methods. So, the AI and machine learning in computational modeling approaches are our extension of traditional statistical modeling approaches. And statistical modeling approaches are able to provide much greater specificity. So, for example, whereas traditional
statistical methods will provide information about a change in X associated with a change in Y, these advanced analytics approaches provide insights not only that a change in X is associated with a change in Y but the magnitude and the timing of that change. So, it allows for greater understanding of those complex and dynamic systems that influence health and health outcomes. Next slide. So, this next slide is another way of looking at it. This is the spectrum of advanced analytics, and it’s a summary of the various types of research questions or health questions that prompts the different types of advanced analytics. So, from descriptive analytics that will inform
questions around, “What has happened to a similar population?” or “What are the trends?” to similarity analytics that informs questions about how to identify best or promising interventions in similar patients, or how to have similar patients really…in...the outcomes with a particular intervention? Or to look at predictive analytics that informs questions about…around what will happen. And finally, prescriptive analytics that informs questions about what should happen. So, again, these advanced analytics approaches are increasingly dynamic, meaning they leverage the dynamic interplay of interactions, observations, and health outcomes over time. And so, interventions are social interactions; the choices that we make, including treatment, being virile; what we observe from the data—all of it can inform decisions or future choices or these analytics. So, they are data points that help us to make better…have better insights that can complement our decisions and interventions in communities and…and environment. But I want to mention that we can improve the accuracy of…of these technologies if we have complete data, right? So, data that’s diverse and comprehensive, that include social factors that we know influence health and health outcomes, especially outside of clinical care. Next slide.
So, I just wanted…these are examples. We’re able to answer all kinds of questions, and these are real questions that have been from the literature or have been modeled with…with machine learning and AI, such as, “How do I optimize care interventions for different populations?” “What are some triggers for disease onset?” So, really, a lot of potential for AI and machine learning and research to provide insights into issues that are relevant in a real-world context of understanding, identifying, promoting better health, clinical care, public health management. The next slide, I do want to provide some examples. I want to illustrate some examples of how we’re using it. So, this is an example of…if you want to ask, like, “What’s my patient’s risk…or a patient’s risk of developing condition X—diabetes or whatever?" And so, this question can be approached by a combination of several features.
So, there’s feature engineering or feature selection methods that will extract those salient or important and relevant features from real-world health data. Predictive modeling that can provide information from cohorts or population groups to address some challenges of data messiness or specificity in our health care data. We can use personalized predictive modeling methods that do not focus on a single global model trained using all the data but rather patient-specific local models to multitask learning methods that simultaneously can predict multiple…risk and could also leverage across risk information and association. So, the figure on the right illustrates the steady setup for a retrospective prediction task using longitudinal patient data.
The…and you can see the diagnosis dates will be the time when the patient was diagnosed with the disease, and then you can look at…a prediction window to define how much time before the disease was diagnosed, and within that window, you want to compute the risk of the disease. So, it could be looking at 6 months, what did they have...1 year or 2 years, depending on the application use case. And then the observation window defines sort of the period of time before index case.
And the data is typically aggregated using those feature engineering methods into a feature vector or matrix representation for downstream modeling. So, this is where it’s really important, also, to acquire expert knowledge about the process being modeled, collecting the appropriate data to answer the desired question; understanding the inherent variation in responses between different population cohorts; and taking steps, if possible, to minimize this variation so that…which may not be apparent. So…and…you know, collecting the right predictors—social, clinical, behavioral—that are relevant for that disease condition and utilizing a range of model types so that you have the best chance of uncovering those relationships among the predictors and their response. Next slide. You can also look at—yes,
this—what has happened to patients similar, right? So, for example, you have...a treatment pathway, and you want to determine precision cohorts of patients that are similar to a target patient that you have, and so this drives…helps drive research in looking at patient similarity methods, looking at temporal event sequence mining methods so that you can identify salient patterns in the data, looking at disease progression modeling to better understand and characterize how a disease evolves over time and how observations in the data are associated with some of these changes. You can look at pathway visualization methods to represent and survey some of these underlying insights. And the figure on the right really shows that, sort of, visualization of that disease progression model trained on patient EMR. And these are different states of disease, the duration, statistics of each stage, the various observation associated and sort of the transition frequencies of the different states. Next slide, I do want to just…I think just to illustrate, this is a case example that I really like using. It’s for Crohn disease—Crohn’s disease—that...we
know that there are significantly…so, this is a treatment pathway, right, so it represents different cohorts of patients that were put on different medications, and we know that significantly fewer patients from this analysis have included…you know, the treatment pathways included biologic therapies compared with non-biologic therapies. Very few patients were even ever initiated on biologic therapy, but we know that there’s been significant progress made in treatments, and we know biologics are our most effective medication, but it’s only being used in a small proportion of patients, suggesting that there are barriers to, you know, prevent optimized patient management. The next example I actually pulled out this. Really interesting from…again, from our MarketScan Database, but this is a treatment pathway for depression. These are called sunburst of treatment patterns, and it starts with first-line therapy, which is the innermost donut to the fourth line, or outer slices, right? And so, each color represents distinct treatment classes, and each layer represents a new treatment outline. And you…so you can see patients who had had a clinical diagnosis of depression—more than twice were in this cohort—and had inpatient visit for depression. And you could see…so, there’s a commercial…those with commercial insurance,
Medicare insurance, and Medicaid insurance. The proportion of patients that do not receive any pharmacotreatment during follow-up ranged from 29 percent to 52 percent. As you could see in this analysis, SSRIs were most common first-line treatment, but however, if you could look…if you see in the Medicaid cohorts, many patients received a sedative or anxiolytic prior to any antidepressant treatment, even though they had a clinical diagnosis. And so, there were lots of patients that…you know, females accounted for 62 percent of the patient population, and there were comorbid conditions.
But the study showed that while general trends across these populations were relatively similar with some important differences, patients that were covered by Medicaid tend to have treatment patterns that were different than the other three groups. More than half of those patients were untreated, and first-line SSRIs used were lower. And so, it represents a population, you know, with high burden of disease, but it appears that they’re getting different…care when it comes to depression compared with other patient populations. And so, this is sort of some insight into how…and you could, you know, slice this by race, ethnicity, geography to really provide those insights. The next slide is…just wanted to share another visual about how patient similarity
networks are being used for precision medicine, and the goal is leveraging all data sources, including rich genomics, biological interpretable -omics data, and this all requires computational methods to support these heterogenous data and to have more…actually, to predict performance. So, just a few examples, and next slide just talks about, you know, current applications hospitals, health systems, public health, everyone is, you know, using to some extent. These are various applications, both AI and machine learning use. So, there’s application in preclinical research,
in population health and public health for tracking epidemics, clinical pathways for informing treatment protocols. We also use it…being widely used in interpretation of medical images—for example, in diabetic retinopathy, screening mammography. And then there are patient-facing applications, such as virtual intelligent agents, that we’re seeing, and it’s also being used to optimize health care or cancer care delivery processes, such as the procurement of cancer drugs, right? Making sure chemotherapy drugs are being delivered to the right place at the right time and...and such logistics, so really widely used applications. So, I want to transition now and talk a little bit about: What are the opportunities for social and behavioral determinants of health? And we know that there are multiple determinants that shape our health, multiple factors—environmental, economic, education—and these socioeconomic and demographic characteristics really have a complex and integrate…integrated relationship with patient risk and vulnerabilities and health outcomes, and they also have a greater influence on a patient’s health, these social determinants, regardless of age, regardless of gender and ethnicity. So, we have a huge opportunity here to think about the social determinants of health data.
Next slide. But we also know that the...you know, these relationships interact across complex and dynamic pathways to produce the health and disparities that we see at the population level. And in some…most instances, the exposures at the environmental level or the neighborhood level may have a greater influence on population health than individual vulnerabilities, although we know, at the individual level, there may be some personal characteristics, including genetic predisposition, that interact with the environment to produce disease. But as a result of this complexity,
there are limitations, as I mentioned earlier, to our traditional epidemiological and statistical adjustment models in handling the computational complexity of these multiple social, demographic, and environmental interactions that are involved in disease risk, that are involved in progression, that are involved in maintaining health and wellness and…and clinical outcomes. Next slide. We also know place matters. We know that where we live can determine how well we live, and it’s a significant factor of health…healthy life and healthy life expectancy. We know food insecurity is a risk factor, and so really, having, you know, knowledge about place and…and community social capital, as well, is also important. In the next slide, I just wanted to share, we’ve done some work around leveraging huge social determinants of health data, but I did…you know, for hospital systems, but I picked this one that I just wanted to share, which may be relevant, and this…this was a study that we conducted to really understand the influence of population-level demographics and social determinants of health on mortality from COVID-19. So, we…we introduced predictive algorithmic modeling and machine learning approaches to study the interactions of those complex, multiple determinants of health at the population level. We…so,
next slide…there were several phases of the methodology that included identifying all the publicly available data and features for the study, so looking at select variables that had complete data across all counties. So, we were looking at the county level, as well. Classification and selection of relevant variables that had strong associations with mortality, and then we did a correlation analysis of the variables in the county clusters in an algorithmic clustering process, because we wanted to look at counties with similar geographic, demographic, and health prevalence status to find out how…whether they were…the COVID-19 mortality and…for comparison and see which social determinant of health had a higher…or most high correlation. Next slide. As part of the methods, we looked at a range of social, economic, environmental, disease-risk prevalence, health care determinants that are known to influence susceptibility to disease and health outcomes. So, these are U.S. Census population estimates. We used the CDC Area Depravation Index, CDC Social Vulnerability Index, and COVID-19-related death counts were retrieved from the CDC’s data reporting from the beginning of the pandemic. So, we initially looked at the
accelerated phase to see which social determinants of health were strongly influencing it. So, in the next slide, what we did…this is actually a map of the U.S. showing the clusters, right, so which county clusters are similar. And 28 of the 34 variables—so we looked at…34 variables—were used as features for clustering. And you could see that, based on the…those…we did
an overselection to help identify those most variant features that impacted mortality out of the…all the variables, and the features are listed here. And based on these features, we…we could see, like, six optimal demographic and socioeconomically distinct county-level clusters. The next slide just…you know…sorry if it’s a lot of text, but it’s just to demonstrate that there were six clusters, and, for example, cluster 4 had 356 counties, and most of these were the Southern Black Belt, the North Slope counties in Alaska, Pine Ridge and Rosebud Reservations, and...and cluster 5 was composed about of 223 counties that included New York, Queens, and you could see the prominent social…social features, right, that included residential segregation, preventable hospital rates, median household income, home ownership rates that were prominent sociodemographic features that had strong positive relationship with COVID-19 mortality. Next slide. Just illustrating sort of the mortality comparison
across those county…distinct county clusters. And you could see here, like, cluster 4 and cluster 5 had significantly higher rates of COVID-19 mortality compared to other clusters. So, in the next slide, again, you could see…now you can illustrate, right, with these ranked variable plot figures that those key sociodemographic and related determinants, in terms of the associations and strengths, either positive or negative, with COVID-19 mortality rates. Of course,
we know population density is consistent across all clusters, but when you look in clusters 4, which includes the Southern Black Belt, Black or African American population or HIV prevalence and employment rate were significantly more correlated to mortality than population density, which was, you know, an eye opener, you know, and not surprising, but it was…this is what we found. And in the next slide, you can see sort of a heat map to show their associations and…and correlation. And, again, each county cluster showed those distinct associations with various demographic or social risk factors. But I do want to point out that, you know, in this clustering, correlation is not causation. As you know, it merely quantifies the strength of this linear relationship between two variables. So, a weak correlation or lack of correlation to COVID-19 mortality may not necessarily mean that that selected factor’s not related to mortality rate.
There might be a nonlinear relationship that requires further investigation, but at least there were…you know, you could identify those factors that needed intervention or needed to be…some investment in the…so, next slide, I mean, basically, you know, you could see that machine learning algorithms can help capture. Using this combination, we can cluster counties by similar demographics and social determinants of health data, disease prevalence and risk, and, as I said, this cannot be done with traditional epi modeling because it’s hypothesis and theory driven, whereas predictive modeling is a whole lot of data driven. I want to really, before—next slide—talk a little bit about…spend some time on AI bias mitigation in our efforts...and because I know this is where there are really several opportunities for social
and behavioral health determinants in the domain, and we often talk about bias, but we look more at the algorithm bias. AI bias is a general concept that refers to the fact that an AI system has been designed, whether intentionally or not, in a way that may make that system’s decision use unfair. And so, sometimes we talk about labeling and modeling and measurement bias or missing, but…but I think that I wanted to share, like, in the next slide, the “Five E’s” of broad aspects of bias and how that can also be linked with AI and how we can understand that bias is everywhere and that we’re all part of addressing bias and how we can really help to mitigate it.
And…I explained earlier. So, data sources for informing clinical evidence, for developing guidelines, for building and training algorithms, for clinical decision, largely comes from research. It largely comes from research trials, it comes from…EHR data, from administrative claims data. And so, we all know that data never speaks for itself and that there’s always a human being deciding how data is funded, how data is generated, deciding on what’s collected, how the evidence is translated into practice. There’s always a human being deciding and interpreting and prioritizing data. So, bias introduced into data and subsequently into
health care or management or decision-making is…is really broader than just the algorithms. And one of the aspects I talk about is…the first one is an evidence bias, right? Searching for my research. And a translation bias. We need to…you know, clinical decision is tied to
clinical trials and rigorous scientific randomized controlled trials or real-world evidence data. I think that we need to look at inclusivity and diversity in our research. And we know that the current state of our clinical trials and scientific studies may not always match the demographics of the patient population who are at risk or suffer from the disease condition. And so, promoting diversity and equity and inclusion and standards into our science, especially in behavioral and social…social determinants, is really key. The next is experience, right? So, the health care provider and experience is an integral part of translating that patient data into improved health outcomes. I always think about the provider bias as the elephant in the room because a patient may arrive at a medical facility with symptoms, and it’s the primary care provider or a, you know, clinician who will determine the cause of action.
Their action or inaction based on an examination of the patient, listening to their story, capturing their beliefs is what’s translated into the EHR data, and so that’s really important. Exclusion is one, right, it’s the third “E.” Critical. Missingness. I have environment, that’s the fourth one. So, looking at…life-course exposures, environmental determinants that need to be integrated. And the fifth aspect is data empathy. And this refers to how much empathy,
how much patient values, preferences, or patient-reported outcomes that are integrated into our decision-making. In all this…these are five aspects of bias that we need to think about because they feed into the algorithms; they feed into clinical decision-making. So, I want to conclude with…I know that we often…there is a human part of AI, and in health care, I think, empathy is a reflection of our compassionate care, a reflection of patient-centered care that takes into account the patient’s perspective and circumstances. In order to provide precision and patient-centered care, we have to…we need to understand or have an understanding of why different belief systems, why cultural biases, language, family structures, and a host of other culturally determined factors influence the manner in which people experience illness, how…why they adhere to medical advice or not, how they respond to treatment.
And these differences are real and translate into real differences in outcomes and care. And so, without humanity, our AI tools and our solutions can exaggerate existing racial inequities and other forms of bias. And so, I…I want to just mention that our humanity and our empathy and our compassion are really important aspects of patient-centered care, important aspects of health equity. And it’s critical because it all feeds into the data that’s used for these AI models and algorithms. My final slide is just an extension of it. I think we have an opportunity in social and behavioral science space to really look at our data—our data sources are currently siloed—and really improve and capture those relevant social, behavioral factors that influence health. Machines have endless capacity to study and identify patterns, streamline data, predict patterns, but I think the interaction between us researchers, health professionals, remains critical because we can provide that empathy and that care and that…for all our patients. So,
people are the recipients of care or health. And so, they really are at the center of health care. And it’s just our high-tech and AI solutions that are part of the solution and really function at the service of humanity, rather than the way around. So, thank you so much for your time and for listening to this presentation. I’m happy to take any questions that you have. Thank you. DR. BETH JAWORSKI:
Thank you so much, Dr. Dankwa-Mullan. That was a phenomenal talk. I really enjoyed it. We had…over 300 attendees who were also able to listen in. We have a number of questions. I don’t…I don’t think that we’ll be able to get to all of them, but in the remaining time, I would love to be able to ask you…ask you just a few. One of them being: Over the past 6 years, you’ve been working at the intersection of health, health equity, and technology. How do…would you say that working at a technology company has added to your already deep knowledge of health care and public health? DR. IRENE DANKWA-MULLAN: Great, great question. Thank you. I mean, I…I feel extremely lucky to be working at this intersection. What is…you know, to help make health care more efficient or leverage
these…data to surface more intelligent insights. And I do work with a multidisciplinary team, but in terms of insights, I think technology and data—big data—has tremendous potential. I don’t think we’re using it to the level that it…its potential. There is so much promise and so much potential to be realized…the health care sector has generated huge amounts of health data. I mean, driven by accumulated biomedical research data, public health data, hospital data. They’re all
meaningful, but they’re so large and so complex to be handled by a traditional software system. And so, we really need the insights…like, we need to get this as a foundation, right, across all of our partners and stakeholders to bring such technology and advanced analytics to the forefront so that we’re able to sufficiently address the complexity of health care. So, I think our science needs to advance in this space. We’ve reached the limits of our human capacity and with manual tasks and with analytics. And we have emerging data now, which is too numerous to compete, and I…I’m willing to…you know, I’m hoping to see more done in the space for the future of health care. DR. BETH JAWORSKI: Thank you. Thank you so much. I think a related question that also touches on some
of the questions that were posed by the audience is about: How do you work to ensure that diverse perspectives and underrepresented communities and their data are included in conversations around AI and ML technology innovations? There were a number of questions that were submitted about that issue of data “missingness” and how that impacts everything across the…the pipeline of…of the work that you’re doing. Would you be able to say just a little bit about that? DR. IRENE DANKWA-MULLAN: Yes, that’s a…that’s a great question, and I…and I do agree, it’s a huge challenge and an issue that we really need to make concerted efforts and commitment to work with communities, work with those, you know, leaders in those communities, bring them to the table, make a conscious effort to be inclusive of their perspective, respect their data. I mean, part of AI ethics is the data is not for the technology company; it’s for the owners or those that generate it. And have some standards or principles, right, or, you know, around how we can work with them on their data. And so, I think we start…need to start
building relationships on trust and transparency and listening and understanding their needs and their values. There’s a whole lot of trust that needs to be done, but really including them at the table in a participatory manner and…and working so that the technologies benefit them. I think that’s where we need…we need to really work on and include them in conversations. I think…I know that at NIH, there is a lot that’s being done around research and diversity and including communities. We have a long way to go, but at least we’re…we’re making…headway with that.
DR. BETH JAWORSKI: And a follow-up question to…to what you just said related to some questions that came in and something you mentioned early on in your talk with respect to language and, I think, the power of language. I’m wondering if you could say a little bit about some of the strategies or the ways that you’ve been able to actually change the language that’s used. I know you referenced some language earlier in software development that used to be very commonly used, and it sounds like you had some success in being able to have folks stop…stop using that discriminatory language. I’m wondering if you could speak a little bit to strategies for changing language and I think, in turn, building…building trust among… DR. IRENE DANKWA-MULLAN: Yeah. Yeah. Thank you for that question. I mean, it’s an ongoing process. We started by…I mean, sometimes people may not realize
certain language will be offensive, so there’s a lot of education and training and…and, you know, you look at the terms that we’re using and ask, “Why is this being used?” Right? And the meaning around it. And so, we started to collect—you know, this was at IBM—to really collect a huge database of…of terms that are used in technology that may have not been appropriate or right. You know, there may be a group of people that already knew these were offensive language or not inclusive or stereotypes, right? So, there are a whole lot of…there’s a whole lot of language out there, but you start with, really, education and starting to build that database. And then, looking at these words, and saying, “Maybe we could replace it with more inclusive.” You know, instead of “master” and “slave,” we have, you know, other terms that will be used. And so, I think it can be done with collaboration. DR. BETH JAWORSKI: Thank you. Thank you. And I think…we may have time for…for one more question.
I would…would love to end with a question about advice. I know that this is a very, very popular area. Certainly, the pandemic has ushered in a wave of digital health technologies. What advice would you give to researchers—particularly folks who’ve been trained in the social and behavioral sciences—at various career stages who are interested in pursuing a career that’s specifically at the intersection of AI and ML and health equity? What…what advice would you have for folks that would like to pursue this career path? DR. IRENE DANKWA-MULLAN: Yes. I think the…I think it’s great. I mean, I…I always say that everyone…every training curriculum needs to have a foundation in AI and machine learning or data science as foundational, as part of the core requirement, because it’s really important. The field is becoming more multidisciplinary, so it’s not now just...you know, science…I mean, math or computer engineering or software or biomedical engineering students that are going into those fields. Our team had included anthropologists and epidemiologists,
public health experts. And there are psychologists or linguistics that are going to…so, I will say, taking courses, taking classes, but it’s very important in your science career or…to look for some of these classes. They are Coursera classes. There are…IBM has a, you know, AI courses or machine learning foundational…foundational courses that you could use. And it’s really a discipline that has…will benefit from rich perspective. So…and really excellent career opportunities, especially in…in this space. DR. BETH JAWORSKI: Fantastic.
DR. IRENE DANKWA-MULLAN: So, I would say go for it. You know, don’t be afraid. Let the…everyone needs to have…take a class in that, and I will be cheering them on. We need more…more early investigators or researchers in this space. DR. BETH JAWORSKI: Wonderful. Thank you so much, Dr. Dankwa-Mullan for that fantastic presentation. Dr. Hunter, I would like to turn it back over to you for closing remarks.
DR. CHRISTINE HUNTER: Thank you. So, thank you, Dr. Jaworski for moderating the question-and-answer session, and a big thank you to you, Dr. Dankwa-Mullan, for your excellent presentation. And also, thank you to the large crowd that joined us online today. If you have colleagues who were unable to join and may be interested in this topic, please remind them that a recording of today’s webinar will be available in about 1 month on the OBSSR website at obssr.od.nih.gov. The next OBSSR Director’s Webinar will be held on September 27 at 2 p.m. Eastern and will feature Dr. Emily Falk, who will present on how health messages
can affect behaviors, such as alcohol and tobacco use. Please also subscribe to OBSSR’s listserv to receive updates on upcoming events. Again, you can sign up at obssr.od.nih.gov. And with that, we conclude today’s webinar. Thank you for attending, and I hope you all have a great day.
2022-08-16 10:25
