Fascinated by video games and puzzles as a child, Marzyeh Ghassemi was also fascinated by health from an early age. Fortunately, she found a way where she could combine these two interests.
“Although I was considering a career in health care, my interest in computer science and engineering was stronger,” says Ghassemi, an associate professor in MIT’s Department of Electrical Engineering and Computer Science and the Institute of Engineering and Medical Sciences (IMES) and principal investigator of the Laboratory for Information and Decision Systems (LIDS). “When I discovered that computer science broadly, and artificial intelligence/learning in particular, could be applied to healthcare, it was an alignment of interests.”
Currently, Ghassemi and her Vigorous ML research group at LIDS are working on in-depth research on how machine learning (ML) can be improved and then applied to improve health safety and equity.
Growing up in Texas and Fresh Mexico in an engineering-oriented Iranian-American family, Ghassemi had role models in STEM careers. Although she loved puzzle-based video games – “Solving puzzles to unlock the next levels or advance further was a very appealing challenge” – her mother also engaged her in more advanced mathematics early on, encouraging her to see mathematics as more than arithmetic.
“Addition and multiplication are basic skills that are emphasized for a reason, but focusing on them can obscure the impression that much of higher-level math and science is more about logic and puzzles,” Ghassemi says. “Thanks to my mom’s encouragement, I knew interesting things were ahead of me.”
Ghassemi claims that apart from her mother, many other people supported her intellectual development. When she earned her bachelor’s degree at Fresh Mexico State University, the director of the Honors College and former Marshall Scholar—Jason Ackelson, now senior adviser to the U.S. Department of Homeland Security—helped her apply for a Marshall Scholarship, which landed her at the University of Oxford, where in 2011 she earned her master’s degree and first became interested in the novel, rapidly developing field of machine learning. During her doctoral work at MIT, Ghassemi says she received support “from both professors and peers,” adding, “It’s an environment of openness and acceptance that I strive to provide for my students.”
While working on her PhD, Ghassemi also came across the first hint that errors in health data could be hiding in machine learning models.
She trained models to predict outcomes based on health data, “and the mindset at the time was to use all the data available. In the case of image neural networks, we have seen that appropriate features can be learned to achieve good performance, eliminating the need to manually design specific features.”
During a meeting with Leo Celi, principal investigator in MIT’s Computational Physiology Laboratory and IMES and a member of Ghassemi’s thesis committee, Celi asked whether Ghassemi had tested how well the models performed with patients of different gender, insurance type, and independence. reported races.
Ghassemi checked and there were gaps. “We now have almost a decade of work showing that these model gaps are difficult to address – they result from existing biases in health data and default technical practices. If you don’t think about them carefully, models will naively replicate and extend biases,” he says.
Ghassemi has been researching these issues ever since.
Her favorite breakthrough in her work so far came at several moments. First, she and her research group have shown that learning models can recognize a patient’s race from medical images such as chest X-rays, which radiologists cannot do. The group then found that models optimized for “average” results didn’t perform as well for women and minorities. Last summer, her group combined these findings to show that the more a model learns to predict a patient’s race or gender based on a medical image, the greater the difference in outcomes for subgroups within those demographics. Ghassemi and her team found that the problem could be mitigated by training the model for demographic differences, rather than focusing on overall average performance – but this process must be performed at each facility where the model is implemented.
“We emphasize that models trained to optimize outcomes (balancing overall efficiency with the smallest equity gap) in one hospital are not optimal in other settings. This has a significant impact on how we develop models for human use,” says Ghassemi. “One hospital may have the resources to train a model and then be able to demonstrate that it performs well, perhaps even under certain fairness constraints. However, our research shows that these performance guarantees do not apply in the new settings. A model that is well-balanced in one setting may not work effectively in another environment. This has implications for the usefulness of models in practice and it is critical that we work to address this issue for those developing and implementing models.”
Ghassemi’s activities stem from her identity.
“I am clearly a Muslim and a mother – both of which have helped shape the way I view the world, which informs my research interests,” she says. “I work on the robustness of machine learning models and how a lack of robustness can combine with existing biases. This interest is not a coincidence.”
As for her mindset, Ghassemi says inspiration often comes when she’s outdoors – biking in Fresh Mexico as a student, rowing in Oxford, running as a graduate student at MIT, and currently walking on the Cambridge Esplanade. She also says that in approaching a convoluted problem, she found it helpful to think about the parts of a larger problem and try to understand how her assumptions about each part might be wrong.
“In my experience, the most limiting factor when it comes to new solutions is what you think you know,” he says. “Sometimes it’s hard to let go of your own (partial) knowledge of something until you dive deep into a model, system, etc., and realize that you didn’t understand the subpart correctly or fully.”
While Ghassemi is passionate about his work, he intentionally follows a broader perspective of life.
“When you love your research, it can be hard to keep it from becoming your identity — something that many scientists should be aware of,” he says. “I try to make sure I have interests (and knowledge) beyond my technical knowledge.
“One of the best ways to prioritize balance is to work with good people. If you have family, friends or co-workers who encourage you to be a whole person, stick with them!”
Having won numerous awards and much recognition for his work encompassing his two early passions – computer science and health – Ghassemi believes in seeing life as a journey.
“There is a quote by the Persian poet Rumi that translates to: ‘You are what you are looking for,’” he says. “At every stage of life, you have to reinvest in finding who you are and pushing that towards who you want to be.”