A file photo of young Tina with a horse

Not her first rodeo

How do you fix AI’s understanding of marginalized populations? Tina Hernandez-Boussard thinks the answer lies in her rural roots.

Story tags:

Tina today

Tina Hernandez-Boussard. Photo: Jess Alvarenga

Share this story

To understand how Tina Hernandez-Boussard is changing the future of artificial intelligence in medicine, you have to start with barrel racing.

The Stanford professor of medicine and biomedical data science grew up a 4H kid in Bishop, California, raising livestock and riding rodeo. She competed in barrel racing—riding horses fast around barrels—and calf-riding. You can think of the latter, which involves strapping children to the backs of calves, as bull-riding for the elementary-school set.

“There’s a rope around the calf and they tuck your hands under it,” Hernandez-Boussard explains. “We didn’t have stirrups so my dad tied my boots to the calf, too, so I couldn’t fall off.”

Hernandez-Boussard loved rodeo and being around family, but she wanted a broader life. She found one through STEM classes in high school. “Data was always kind of my thing,” she says. “Calculus, computer science, it was easy for me.”

Nobody in Hernandez-Boussard’s extended family had gone to college. Her father was an agricultural worker and, like a lot of rural families, they never had medical insurance. Hernandez-Boussard recalls only two doctor’s visits in her entire childhood: once for a broken shoulder, once for a vaccination. Otherwise, she’d see neighbors, friends, or veterinarians for medical help. It was a small town and the community took care of itself.

A file photo of young Tina with a sheep at the rodeo

Long before there was biomedical data science in Hernandez-Boussard’s life, there was livestock and rodeo. Photo courtesy: Tina Hernandez-Boussard

But she got into UC Irvine and then, after graduation, a master’s program in public health at Yale. She worked next as an epidemiologist in Europe, got a PhD in computational biology and in 1999 joined Stanford as a staff scientist, building databases for gene-expression analysis. Later, she joined the surgery department at Stanford School of Medicine to create computational tools using clinical data. Every year seemed to take her further from Bishop—though Bishop would ultimately come full circle in her work. 

Hernandez-Boussard was one of the first data scientists using electronic health care records to train computer models—a kind of precursor to contemporary AI. The resulting computational tools, she explains, can be applied to complex clinical issues like pain management by learning from retrospective data for patient demographics, diagnoses, previous treatments, and previous utilization of health care resources. From there they can find patterns in patient response to pain medication and determine a course of care.

Now, in the age of generative artificial intelligence, Hernandez-Boussard sees the potential for computational medical tools of vastly greater power. Where once she might have trained a model on data from 300 patients, Hernandez-Boussard says, she’s now using large language models trained on billions of parameters. 

“Before, I could predict something for a given patient and say, ‘There’s a 70 percent chance we’ll have a bad outcome,’” she says. “Now I can say, ‘I’m 96 percent sure.’” 

But a stark truth has also become increasingly clear. With great accuracy—to paraphrase Spiderman—comes great responsibility.

Strung between Hernandez-Boussard’s rural upbringing and her work at Stanford is a simple fact: The more we trust an AI tool to be accurate, the more likely we are to follow its recommendations. That might be fine with an individual patient who bears a close resemblance to the average patient in the training data. It becomes dangerous when the individual patient, like a lot of the people Hernandez-Boussard grew up with in Bishop, have little in common with the average patient in the AI tool’s original training data.

Before, I could predict something
for a given patient and say, ‘There’s a 70 percent chance we’ll have a bad outcome.’ Now I can say, ‘I’m 96 percent sure.’”
Tina Hernandez-Boussard
Michael McFaul sits at a desk chatting with student Andrii Torchylo

Photo: Jess Alvarenga. Background photo courtesy: Tina Hernandez-Boussard.

“They don’t trust the system”

Take, for example, an AI tool that uses racial identity as a parameter. Tell the AI that a patient is Latinx, or speaks a language other than English, and the AI will factor that into its risk prediction; but the original training data included so few Latinx patients that this data point can have outsized and unpredictable implications.  

At the root of this problem, Hernandez-Boussard says, is the fact that race is a social construct without any firm genetic basis. As she puts it, “There’s more diversity within a race than there are between races.”

At the same time, our socially constructed racial identities have real impact on people’s lives, such that social determinants of health, like income and education, tend to cluster around them. So, when an AI model produces medical-treatment outcome probabilities based in part on racial identity, it is really using racial identity as a vague and unreliable proxy for something that hasn’t actually been measured—and for which there is no data.

To illustrate this point, Hernandez-Boussard describes an AI-driven risk calculator for adverse outcomes in Cesarean birth among women who have previously given birth vaginally. This calculator included being Black as a risk factor—even though it wasn’t. The real risk factor, for which race merely served as a proxy, was chronic hypertension. Chronic hypertension data hadn’t been collected, though. Race had.

And so race, instead of hypertension, became the parameter. This meant that the calculator might generate an inaccurate—and excessively high—risk of complications from Cesarean delivery for a Black mother who did not have hypertension. It might also generate a misleadingly low risk for a white mother who did have hypertension, putting both women in danger.

There is a very real concern, Hernandez-Boussard explains, that such problems will worsen as AI moves further into health care—especially in rural areas with marginalized populations like Bishop. These areas often operate as sort of “health care deserts,” which put huge burdens on the providers who do serve them. Hernandez-Boussard worries that, in that context, the allure of AI to save time and money will mean it’s deployed for populations most likely to be underrepresented in the training data—and, therefore, most likely to receive harmfully inaccurate diagnoses and outcome predictions. 

That means that one of the great social justice challenges in AI research lies in Hernandez-Boussard’s field. Researchers have long known that medical data sets come overwhelmingly from the white majority population most likely to fill out questionnaires. Previous efforts to correct this have used the same data-collection methods—and failed.

“They don’t trust the system,” Hernandez-Boussard says, explaining why people in marginalized populations resist sharing data. “We’ve harmed these populations many, many times.” She cited the infamous Tuskegee Trials of the mid-20th century, when the U.S. Public Health Service studied 399 Black men with syphilis by deliberately not treating them, despite the availability of effective medication.

They don't trust the system.
“We’ve harmed these populations many, many times.”
Tina Hernandez-Boussard
Michael McFaul sits at a desk chatting with student Andrii Torchylo

Photo: Jess Alvarenga. Background photo courtesy: Tina Hernandez-Boussard.

What it takes

Turning things around demands a new approach—for which Hernandez-Boussard is uniquely qualified. In her latest role as associate dean of research for the School of Medicine, Hernandez-Boussard focuses on educating faculty across the university in the equitable application of AI. She works with assorted teams at the School of Medicine to ensure appropriate application of existing AI diagnostics for patients from marginalized populations. Finally, Hernandez-Boussard has become a thought leader in the national conversation about equitable AI, working with diverse stakeholders to develop standards and guidelines in the future development of AI healthcare tools.

Throughout this work, Hernandez-Boussard draws frequently on lessons learned in Bishop.  

We, as scientists, expect patients to
go out of their comfort zone and come to meet us. We’re the ones with resources. Why can’t we go out of our comfort zone?”
Tina says about traditional approaches to data collection.

Hernandez-Boussard cites research showing that the process of collecting data from marginalized populations often works best in places like barbershops and churches, where people naturally congregate and feel comfortable. That has her thinking of her hometown—places like the Bishop Pauite and Shoshone Indian Reservation, restaurants like Whiskey Creek and El Charro, or the city park. 

Most of all, when Hernandez-Boussard thinks about gathering data in rural communities like her own, she thinks about a big annual tradition in Bishop called Mule Days, held every summer with country music bands, rodeo, and more than 700 mules and trainers competing in cattle working, mule jumping, and chariot racing. 

“Mule Days is something I would go to,” she says. “Because everyone comes out and there’s Indian tacos—that’s the big thing in Bishop, like Indian fry bread made into a taco. You get out to a place like that, in person, and make those human connections.”

Hernandez-Boussard has already begun. “We did a program trying to engage with Native Americans,” she says, “so I just called one of my girlfriends at the Indian Health Center. I'm like, ‘Want to do this with me?’ And she said, ‘Absolutely!’ That's the way to get engaged, knowing somebody who will let you into the community—rather than just sending out emails that say, ‘Come down to my office and I’ll give you 25 dollars to take a survey.’ They’re never going to do that. You have to show that you want more than just their data and you're going to do good with it.”

The stakes are high, says Hernandez-Boussard, when it comes to the equitable engineering of health-care AI.

“If we get this right,” she says, “some of these underrepresented populations can benefit the most. We can really close the racial gap in health outcomes. If we don’t, we’re widening that gap. It's all about building the trust to make responsible use of AI.”

Learn more

RAISE Health, a joint initiative between Stanford Medicine and the Stanford Institute for Human-Centered Artificial Intelligence (HAI), hosted its first convening on May 14.  You can watch recordings here. To learn more about RAISE Health, visit the website and read its newsletter.

Share this story

Related stories:Explore more