How AI can be used to promote public and population health

TL;DR

This paper explores how Microsoft's AI for Good Lab uses artificial intelligence to improve public and population health, focusing on maternal, fetal, infant, and rural health.

Contribution

The paper presents practical applications of AI in public health and highlights lessons learned from cross-sector collaborations.

Findings

AI can improve maternal, fetal, and infant health outcomes.

Collaboration across sectors enhances AI's impact on population health.

Focusing on health metrics, not just model accuracy, is crucial for real-world impact.

Abstract

Here, we summarize the work that Microsoft's philanthropic Artificial Intelligence (AI) for Good Lab has completed in the realm of promoting public and population health. In particular, after providing examples of how the AI for Good Lab has articulated the value of using AI to improve public and population health, we provide examples and references of the work demonstrating how the Lab has: applied Artificial Intelligence (AI) to improve maternal, fetal, and infant health; leveraged large language models to improve population health; and applied AI to improve rural health and healthcare. We also summarize what we have learned through our work, finding that: getting the question right and ensuring the limitations of any analysis are understood is important; collaboration across public, private, and educational institutions with subject matter experts will be the most effective and efficient way to harness this new technology; and that focusing on metrics that reflect health, and not just the accuracy of the model, is the most impactful way to improve the health of populations, worldwide.

Full text

Introduction

Healthcare delivery and public health efforts seek to improve the health of the population served. While healthcare delivery is an important aspect of population health, public health is even more so: only about 20% of a population's health status can be attributed to health care delivery, with the remaining 80% attributed to genetics and public health measures that include health behaviors and social determinants of health, defined as the non-medical factors that affect health outcomes. While prior public health research has relied on traditional regression and other analytic methods, using artificial intelligence (AI) to address public health issues offers a relatively new approach that is able to explore non-linear and interactive relationships more adroitly.

Given the complexities of the interactions of healthcare delivery, genetics, health behaviors, and social determinants of health, AI may be able to uncover important insights regarding relationships between measures of public health and health outcomes. Should such relationships be revealed, policymakers wanting to improve population health might focus their efforts on addressing those health behaviors or social determinants that most impact health status. Such geographically- or specified-population-targeted efforts could not only use scarce resources more efficiently but also generate a measurable health-related return on that investment. For example, while research has demonstrated that improving local economic conditions can reduce local morbidity and mortality rates, AI can help model the timing, magnitude, and beneficiaries of those potential returns so that policymakers can make informed investment decisions about which interventions should be undertaken to generate the greatest returns to health over a specified period.

To reveal those relationships, since 2020, Microsoft's philanthropic AI for Good Lab has dedicated resources to exploring how AI can be used to address the biggest problems in sustainability, humanitarian action, and health (1). The Lab includes (generally PhD-level) Microsoft-employed data and computer scientists who collaborate with not-for-profit organizations (like universities, charitable organizations, and government agencies) to apply AI methods to research intent on solving those problems. As a segment of the broader literature examining how to use AI to promote public and population health, this paper shows how this Lab and these researchers have contributed to the peer-reviewed literature addressing health and healthcare challenges in the following five areas.

Articulating the potential uses of AI in public and population health

AI for Health researchers have contributed to the literature by articulating the promise and perils of applying AI methods to improve population health. These contributions include editorials asserting how AI might be used to advance public health (2), why AI should be considered a public service (3), and how AI methods might be used to improve health and public health (4).

In addition, AI for Health researchers have emphasized the centrality and importance of population health in efforts to achieve sustainable development goals (5) while also cautioning that inaccurate storytelling used to persuade public health practitioners and policymakers about the importance and relevance of public health concerns may mislead decision makers (6).

AI for Health researchers have also explored how healthcare delivery might be improved using AI methods. For example, they have described the promise and peril of the use of AI in achieving the quadruple aim in healthcare [enhancing the patient and provider experience, improving and equitably distributing health care quality and outcomes, and reducing per-capita healthcare costs (7)] as well as how AI might help address the underperformance of the United States healthcare system (8).

Using AI to improve maternal, fetal, and infant health

AI for Health researchers have made substantial contributions to the literature regarding sudden unexpected infant death (SUID). SUID is an umbrella term that includes sudden infant death syndrome (SIDS), accidental suffocation and strangulation in bed, and other deaths from unknown causes that occur in infants less than 1 year old; it is a devastating experience for a family.

To try to reduce the incidence of SUID, AI for Health researchers have used AI methods to explore the relationship between maternal smoking before and during pregnancy (9), the relationship between maternal infections in pregnancy (10), and the relationship between maternal obesity (11) and the risk of SUID. Further, AI for Health researchers have explored geographic variation in SUID within the United States (12) and identified distinct populations of SUID based on the age of the infant at death (13), including an analysis of SUID in the first week of life (14). Finally, AI for Health researchers found that rates of SUID changed during the COVID pandemic (15).

Beyond SUID, AI for Health researchers have identified risk factors for late gestational stillbirth in the United States (16), worked to improve the understanding of the relationship between fetal growth and late gestational stillbirth (17), and identified risk factors for postpartum hemorrhage and severe maternal morbidity in low-risk laboring populations (18).

Finally, AI for Health researchers have described how AI can be helpful: in addressing challenges in maternal, newborn, and child health in low resource settings so that policymakers can best use scarce resources (19); understanding how micro-geographic analysis can be used to identify locations where children experience high malnutrition risk in India so policymakers can target interventions to those who most need them (20); determining how AI—in concert with satellite data—can be used to anticipate surges in childhood malnutrition in Kenya so that policymakers can intervene to prevent it (21); and modeling the meningococcal antibody response following vaccination so that policymakers can determine when best to revaccinate (22).

Using AI to explore relationships between social determinants of health and population health

As is well-known, the social determinants of health—the conditions in which people are born, grow, work, live, and age—have a greater impact on one's health than does the healthcare delivery system. Understanding that impact on individual and population health, AI for Health researchers have undertaken research to explore the specific relationships between social determinants of health and population health.

Having considered how social determinants of health might be used to address cardiovascular disease and health equity (23), AI for Health researchers have described the relationship between local economic distress and both life expectancy (24) and inequities in health outcomes, clinical care, health behaviors, and other social determinants of health (25). AI for Health researchers used this knowledge and machine learning tools to evaluate the relationships between social determinants of health and diabetes prevalence in New York City (26) and changes in local economic conditions and drug overdose deaths in the United States between 2000 and 2019 (27). The insights generated from examining these relationships support previous work suggesting that policymakers should consider using local economic stimulus to improve health and health outcomes (28).

Further, AI for Health researchers have explored the Medicare fee-for-service population, identifying discrepancies in coding practices that are meant to capture social determinants of health (29) and identifying where there are shortages of qualified proceduralists necessary to provide high-quality, best-outcomes procedural interventions (30).

AI for Health researchers have also used machine learning, social determinants of health data, and other data sources to forecast short-term drug overdose deaths in Connecticut (31), improve flu tracking (32), predict quality of life trajectories among patients with different types of cancer (33), and explore and challenge inherent assumptions when examining historical data when cultural norms have changed (34).

Using large language models to improve population health

Large Language Models (LLMs) have the potential to dramatically impact healthcare delivery, public health interventions, and health outcomes. To advance this work, AI for Health researchers have demonstrated the efficacy of a conversational chatbot for cigarette smoking cessation (35), articulated how LLMs could be used to facilitate shared-decision making in the physician-patient encounter (36), demonstrated how LLMs can be used to quickly and efficiently create user-friendly applications for assessment of eligibility for Medicaid or other social services (37), to evaluate the accuracy of answers to nutritional questions in Brazilian Portuguese (38), to identify sources of misinformation in internet news stories (39), and to surface high-quality, evidence-based medical information in internet searches (40).

Using AI to improve rural health and healthcare

To conclude a synopsis of the program's efforts examining health and public health, AI for Health researchers have been instrumental in providing a mapping tool and evaluating rural-urban disparities in health outcomes, clinical care, health behaviors, and social determinants of health (41). To address the health and economic disparities this work found, AI for Health researchers developed an action plan to leverage technology and AI to improve the health of those living in rural settings (42) and devised a hub-and-spoke model designed to improve rural health and healthcare access by facilitating clinical and technological partnerships between urban and rural hospitals (43).

What we have learned

Our experiences have generated several insights that might be useful to others exploring the use of AI in public and population health. First, as in all research, it is critically important to get the question right. While “fishing” through data exploration might generate interesting hypotheses, it is critical to understand the analytic approach to be used, to appreciate the limitations of any datasets to be explored, and to obtain the correct permissions—including human subjects review, where appropriate—to conduct the research. Second, and a corollary to the first point, the questions that AI has the potential to answer are challenging and complex: it is therefore imperative to include subject matter experts (in public health and in the particular medical area being explored) in the development of the question, the research itself, the interpretation of findings, and the write-up. Third, to achieve the first two points requires collaboration with institutions that have the appropriate data for analysis and the subject matter expertise to conduct the research; selecting the best partners who are interested in sharing findings, regardless of what they are, in an effort to advance knowledge is paramount to success. Finally, research undertaken must have the potential to be impactful: researchers must apply their limited resources to questions that can improve public and population health; therefore, the metrics that should be evaluated should be reflective of health impact and not only measure of model accuracy.

Conclusion

Here, we have shown how AI can be—and is being—used to promote public and population health. While AI for Health researchers have been incredibly productive in contributing to this area, we believe that this is just the beginning. In the future, we anticipate that AI will be used to integrate individualized healthcare delivery and social determinant of health intervention plans, estimate returns to investment in targeted social determinant of health interventions, use the Internet of Things to surveille for and intervene on infectious disease outbreaks, and use multi-modal data (such as satellite data) for epidemiological purposes, climate-change induced disease spread, and even concurrent evaluation of social determinant of health metrics.

We live in an exciting time. While AI is revolutionizing how many businesses operate, its greatest impact may be on improving public and population health. A comprehensive, strategic deployment of particular AI tools that target public and population health interventions, support administrators in montitoring those interventions, and inform policymakers on which interventions to choose can efficiently and effectively support a learning cycle, wherein individual and population health continuously improve.