Jun 27, 2024
How a not-for-profit health plan is working to remove bias from AI
To ensure members receive the care they need, researchers scrutinize their algorithms for inequities
Massachusetts’ largest health plan is taking cutting-edge steps to address potential bias in artificial intelligence and ensure the technology can be used to enhance access to care for patients.
“In a world of inequity, we can’t just take data at face value,” said Dr. Deborah Peikes, vice president of measurement and evaluation at Blue Cross Blue Shield of Massachusetts. “We have to take a more holistic approach and understand what data we are feeding into our models because there can be unintended consequences where we are further embedding inequities in care.”
Facing the challenge
Artificial intelligence, or AI, is used in many fields, including health care, where it offers the potential for increasingly sophisticated and personalized care. AI uses computer systems, or “algorithms,” to analyze and learn from vast amounts of data, getting “smarter” along the way, like humans do.
However, research has shown algorithms can form biases as they are fed data that reflects racial or other inequities.
For example, a landmark 2019 study showed how an algorithm used in one large hospital system resulted in less follow-up care for Black patients.
The model was designed to evaluate how much money had been spent on each of the patients in its system and to flag high-cost patients as “high-risk” and in potential need of follow-up care.
But, the study noted, the data was skewed by structural inequities in the health care system: “unequal access to care means that we spend less money caring for Black patients than for white patients. The algorithm thus falsely concludes that Black patients are healthier than equally sick white patients.”
“The study taught us that just because a certain population uses less care doesn’t mean they have less need or are less sick,” said Peikes. “It also served as a catalyst for the health care community to continually examine our algorithms.”
In this case, researchers found the inequities could be addressed by modifying the algorithm to evaluate the complexity and severity of patients’ conditions — rather than the cost of the care they were receiving.
Finding solutions
Blue Cross uses AI to improve access to care for patients by providing tailored recommendations for primary care and mental health clinicians, combating health care fraud and ensuring vulnerable members receive needed support.
“Like all health care organizations, we know that sometimes computer-generated models can lead to unintended results, including bias,” said Dr. Mark Friedberg, an internist and senior vice president of performance measurement and improvement at Blue Cross.
The not-for-profit health plan regularly conducts “equity audits” on all its artificial intelligence-driven programs to ensure these models are delivering care in an equitable way among its nearly 3 million members, regardless of their race or ethnicity.
Our goal is to use all of our available data to create an exceptional experience for our members and to help them improve their health. We know that data can contain bias, and we work hard to stay vigilant and remove as much bias as we can from our models.
- Himanshu Arora, chief data and analytics officer at Blue Cross
For example, Friedberg noted, a model may indicate that white patients are sicker than Black and Hispanic patients because white patients use more health care services on average.
“For many years, we have conducted regular reviews of our models to make sure our Black and Hispanic members are not being negatively affected,” he said.
The health plan pays particular attention to algorithms that identify members who might benefit from programs that provide support for conditions such as diabetes, kidney disease and heart disease.
For instance, the plan regularly compares the proportion of Black and Hispanic members its algorithms flag for cardiovascular programs to the overall proportion of Black and Hispanic Americans who have cardiovascular disease to make sure its algorithms aren’t undercounting these members. If the team does identify an issue, it acts quickly to remedy any disparities.
A broader effort
Monitoring and removing bias from its algorithms is just one way Blue Cross is tackling inequities in health care.
The health plan is collecting race, ethnicity and language data from its members and using that to publish an annual Health Equity Report. It also has awarded $25 million in grants to large and small provider groups to help them make the business case to invest in health equity and signed groundbreaking contracts with providers linking financial incentives to achieving measurable improvements in health equity.
This initiative is just another way that we are holding ourselves accountable for making health care more equitable for all of our members.
Dr. Mark Friedberg
Learn more about how Blue Cross has used artificial intelligence and data and analytics to combat health care fraud and reach out to vulnerable members during the pandemic.