Why We Need to Build Diversity Into Our Algorithms
The COVID-19 pandemic put a spotlight on racial, ethnic, and socioeconomic disparities that have plagued healthcare for decades. New CDC research found that Black, Hispanic, and Native American people were more likely to be hospitalized for COVID-19 than white people. In some cases, these groups were 1.7 times more likely to end up in the hospital emergency department with COVID-19. The research points to longstanding, systemic inequities, including limited access to quality healthcare, disproportionate representation in essential jobs, and overall socioeconomic status, as the drivers of the trend.
These are exactly the types of gaps in care that machine learning and artificial intelligence (AI) -driven health tech solutions have promised to close. By analyzing massive amounts of patient data, scrutinizing patterns of behavior, and surfacing anomalies that are consistent with inequity, or uneven care delivery, today’s healthcare analytics algorithms should be able to give us the clues we need to remove bias from the equation.
Just as more industries have started to deploy AI over the past decade, we’ve also seen that even algorithms have bias – and those biases can have harmful results, from racist Twitter chatbots to exam-scoring algorithms that skew against public school students. And in healthcare, when AI is built on biased data, it can actually increase the divide between the “haves” and “have-nots,” rather than closing gaps in care.
How do we grapple with both sides of this complex equation: the incredible role that data and AI can play in supporting better care and equity across healthcare – and the danger of exacerbating gaps in care for patients already most at risk?
Komodo Health co-founder and president Web Sun sat down at MedCity INVEST this week to discuss this issue, joining a panel of health tech leaders to discuss the role of AI in health equity. Read on for more of Web’s insights on how to responsibly, equitably, and powerfully deploy data and AI to help reduce the burden of disease.
Machine learning has promised to enable physicians to deliver more personalized medicine, find and close gaps in care, and increase efficiency in diagnosis and care delivery. Will AI ever be able to live up to these massive promises?
The short answer is ‘yes.’ However, effective algorithms require access to real-time, comprehensive, and connected data. AI is only as good as the data it's built on, and AI innovation needs to be built on a canonical view of patients and all the healthcare actors around them.
Candidly, this is no longer a technology problem, but a problem of data ownership and lack of collective will to get it done. Our healthcare system is deeply entrenched in a culture where each silo operates only within the context of its own proverbial “four walls,” making it difficult to create the deep and broad pool of data to power good algorithms.
How have you seen healthcare stakeholders – whether life sciences companies, payers, advocacy groups, or other organizations you work with – use AI insights to address systemic bias in the system?
As one example, we see payers using AI to access timely insights to better understand underserved and at-risk populations and better meet their needs. The stories associated with health disparities and social determinants of health highlight important opportunities for payers to extend access and address challenges ranging from food and housing insecurity, to “economic dislocation,” and unequal access to care – all of which have been heightened by the pandemic.
We see AI and real-world data helping payers better understand these challenges and update programs to specifically target the needs of underserved patient communities. We even worked with a payer who used geo-coded information along with Komodo’s Healthcare Map™ to pinpoint primary care clinics in underserved communities that were at risk of shutting down. The payers were then able to reach out directly to help those clinics avoid closing down and leaving their communities without care.
That’s powerful. But this kind of impact starts with data that encompasses the depth, breadth, and diversity of patient experiences in our system. A lot of companies that are assessing bias in the system start with commercial data sets because they are often the ones that are most complete. But they also skew healthier and wealthier, and therefore offer an incomplete view of the patient experience.
Help me understand more about the inherent biases that come from disconnected, siloed health data.
Today, most health data comes from systems of record. Each of those systems misses a part of the patient experience. A healthcare organization might not see when the patient goes to a competing hospital, for instance, or when a patient moves out of state. An advocacy group often misses patients who fit into their population and could benefit from their services. A payer doesn’t know when a patient pays out of pocket for a prescription because it’s cheaper than their insurance. The uninsured are often left out of datasets entirely. That lack of visibility into how patients traverse the entire system over time creates a gap and means providers, patients, life sciences companies, payers, advocacy groups, and other stakeholders are making decisions based on an incomplete picture.
Despite years of new healthcare innovations, new medical breakthroughs, and even research that looks deep at health disparities, it seems we still are not making a dent in outcomes for at-risk populations. Where do we start?
To begin to address this problem, we first need to quantify the scope of it, identify at-risk segments of the population, and accurately target interventions that will have the most significant impact. We can’t afford to allow biases in our systems and our data to undermine efforts toward equity.
Look at telemedicine as an example:
Telemedicine was supposed to be the great equalizer. And it was largely seen as one of the silver linings of COVID. But when we applied data on median income to our Healthcare Map, we found that patient utilization of telemedicine has been directly correlated with income during COVID. In fact, we found that providers in the 10% of U.S. counties with the highest median income were 47% more likely to bill for telemedicine than those in the bottom 10% of U.S. counties by median income level. This is a crucial gap for a service that has been a lifeline for many patients during COVID.
COVID mitigation and lock-downs had a significant impact on care for a host of health concerns. What are you seeing with respect to the neglect of chronic diseases, many of which already affected communities of color disproportionately?
We have essentially seen 2020 as a “gap year” for routine screenings. Looking at our data, we’ve found secondary effects of the pandemic that will continue to ripple for years to come. For instance, lipid panel tests – a proven method of spotting cardiovascular risk – were down 65%. A1C tests fell by 61%. Mammogram volumes were down 96%; Colonoscopy volumes were down 91%; Cervical cancer screenings were down 72%. Preventive screenings and care patterns continued to lag behind 2019 volumes throughout the summer of 2020, and have not made up for the ground lost during the height of the pandemic.
These diseases aren’t going anywhere; they are just going unmanaged and undiagnosed. This means they are going to be detected and treated later, or chronic disease will go unmanaged. So how will our health systems cope as latent diseases emerge? We need to keep looking at the data and tracking trends in screenings as well as diagnosis and treatment rates, to better understand the impact of the pandemic on chronic disease management and outcomes as the effects of the pandemic continue to ripple out.
The pandemic has also illuminated the social determinants of health, and particularly racial health disparities. How can we use the learnings of the past year to move forward toward a world where race, socioeconomic status, neighborhood, and so on, don’t make you more susceptible to chronic disease or poor health outcomes?
Unfortunately, the divide between the “haves” and “have-nots” has gotten even worse amid COVID. We see how that has played out with COVID itself, but it ripples out to other health concerns as well.
We can look at diabetes as a case study in how different race and socioeconomic factors layer on top of one another to form really concerning patterns. Diabetes has always disproportionately affected racial/ethnic minority populations. In 2016, the Diabetes Journey said, compared with white adults, the risk of having a diabetes diagnosis is 77% higher among African Americans, 66% higher among Latinos/Hispanics, and 18% higher among Asian Americans.
Now, many of these same groups are also facing higher unemployment rates as a result of the pandemic – meaning many of the same populations who are more likely to face diabetes now are also facing a loss of income and health insurance. Insulin is life-saving but costs an average of $6,000/year – not cheap, especially for patients facing a loss of income or insurance. Black individuals are also less likely to use insulin pumps and other disease management tools. Why is this? Are healthcare professionals failing to write prescriptions for certain populations? Is it a lack of trust in the system?
We need to see and identify these kinds of unmet needs in order to understand and address them.
But comprehensive data is going to be essential to mitigate bias as we answer these questions. To address the unmet needs of entire communities, you have to have access to data that reflects the epidemiology of disease and the experiences of the population. We believe we are finally getting to a point where we can break down these silos, co-mingle data, and reach richer insights and meaningfully address these challenges.