Artificial intelligence (AI) has emerged as a powerful tool for making population health analytics more accurate and interventions more effective. Not all AI-powered pop health tools are created equal, and while many still exhibit different types of bias that limit their ability to accurately identify issues affecting certain populations, we have learned a lot about the sources of bias in AI and how to eliminate them. AI shortcomings aside, pop health analytics have advanced to the point that we are now getting very good at finding groups and individuals with gaps in care. Unfortunately, we still have a lot to learn about how to effectively reach many of those cohorts with appropriate interventions.
Social determinants of health (SDOH) such as housing, food and transportation insecurity; lack of health equity and discrimination; and language and literacy challenges pose significant obstacles both for identifying groups with gaps, and also for addressing those gaps effectively. To be successful, providers and community outreach teams need flexible, data-driven tools to help them design and carry out interventions in ways that are culturally-, community- and individual-appropriate. They can also benefit from building on the best practices and hard-won wisdom developed from effective pop health interventions for SDOH communities in recent years.
AI Promise and Problems
Advances in AI and machine learning have made it an increasingly useful tool in pop health. Sorting through masses of healthcare and related data to find the gaps, patterns and common contributors to health problems is in many ways the perfect job for the sheer computational power of AI. AI is also better at identifying relevant patterns than humans because algorithms can be built to eliminate many of the potentially erroneous preconceptions that people bring to analysis. However, the technology still has some critical shortcomings that must be understood and addressed before we can comfortably rely on AI as a key component in pop health programs.
Bias remains a big concern in AI development and use. Code is written by humans who have unconscious biases, and those biases can creep into the algorithms of AI when Quality Assurance functions are not properly sensitized to detect them. Another form of bias can stem from the fact that AI algorithms are often ‘trained’ on sample data sets to optimize performance and pattern recognition. When those data sets are of insufficient size or are skewed – over/ under-representing certain groups — the AI can develop blind spots which limit its ability to accurately identify issues impacting those groups. Sadly, because healthcare data sets tend to be least complete and robust for disadvantaged communities, this form of bias tends to impact those groups disproportionately.
Sensitivity to a wider range of factors that can impact population health, including SDOH, cultural and community issues, is key to minimizing both coding and data-driven bias. As we continue to develop deeper understanding of the roles these societal issues play in health and in our ability to improve health for individuals and groups, we need to feed those learnings back into AI design and testing.
SDOH and Pop Health Interventions
AI remains an extremely powerful tool, and while population health professionals have become very sophisticated in identifying groups with gaps in care, designing effective interventions to close those gaps is often a much more subtle and complex endeavor. Limitations in available resources often make it impractical or impossible to intervene for certain groups/conditions, and SDOH issues represent the most significant set of challenges to successfully intervening with targeted groups.
Housing insecurity is a classic example here. Individuals who are homeless or itinerant often struggle with multiple chronic co-morbidities. They are also hard to find, hard to communicate with and hard to heal. SDOH for this group is both a contributing factor to disease and an obstacle to effectively closing gaps in treatment and prevention.
Looking at a wider range of SDOH factors, including poverty, housing, transportation, food security, education, addiction, cultural factors, and community issues, we see that these issues are often interrelated and cannot be addressed as stand-alone challenges. There can also be a hierarchy of SDOH where groups affected by more fundamental challenges like housing insecurity need support at that level first before higher-level interventions (e.g. – culturally-appropriate education and communications) can be addressed.
Addressing basic needs like housing is complicated and very resource intensive, and while healthcare organizations need to focus on the care issues they actually have the ability to impact, we cannot just ignore the ones that are beyond our reach, especially when they impact the groups already most at risk and disadvantaged.
Pop Health Interventions and SDOH: Best Practices
In our work supporting healthcare organizations seeking to extend pop health programs into communities impacted by SDOH, we have identified a few best practices that should be top of mind for teams building engagement programs:
1. Thoughtful and collaborative program design. These are complex issues affecting diverse communities, so take your time and consult a multi-disciplinary team.
2. Make sure you have a clear problem statement. Know specifically what issue you are trying to address. Be open to changing it based on what you learn in the field but keep it specific.
3. Keep it specific. Narrow down to a target group with common risk factors and SDOH.
4. Identify individual needs to be addressed. These can be medical, educational or behavioral.
5. Develop a clear hypothesis. Make your best assessment of the causes and effects at work and design your interventions accordingly, but always be willing to test and adjust that hypothesis as you go.
6. Know your available resources. Scope your program to available resources or risk running out of runway before you can achieve meaningful results.
7. Gather community-level data. No battle plan survives first contact with the enemy. Document what you learn as you field your program and use that data to test and refine hypotheses, make the case for additional resources and share with colleagues. Good sources include member assessments, and feet on the street (social workers, community leaders, care teams, etc.).
8. Pilot – Measure – Expand – Repeat.
About David J. Sand & Pravin Pant
David J. Sand, MD, MBA, ZeOmega’s Chief Medical Officer, provides clinical direction on product development, positioning, and relevance for the company’s evolving markets. Dr. Sand is a board-certified otolaryngologist-head & neck surgeon who transitioned from a solo clinical practice to chief medical officer roles serving Medicare Advantage and Medicaid health plans. He has participated on the board of the National Association of Independent Review Organizations (NAIRO) and the American Diabetes Association, and has also been a Fellow of the American College of Surgeons, the American Institute of Healthcare Quality, the Institute of Medicine of Chicago, and the American Academy of Facial Plastic and Reconstructive Surgery.
Pravin Pant leads ZeOmega’s advanced analytics team as Vice President of Advanced Analytics, working on new and existing AI solutions and social determinants of health (SDOH) solutions to create best-in-class software products. Pravin has worked in the hospital & health plan industry for over 20 years, during which he has led initiatives in team building, leadership, innovation, artificial intelligence, big data, predictive analytics, and population health management.
