From a young age, Marzyeh Ghassemi found enchantment in the worlds of video games and puzzles, alongside a keen interest in health. Fortunately, she discovered an exciting opportunity to merge these passions.
“Initially, I contemplated a career in health care, but my interest in computer science and engineering ultimately took precedence,” shares Ghassemi, an associate professor in MIT’s Department of Electrical Engineering and Computer Science as well as the Institute for Medical Engineering and Science (IMES). She also serves as the principal investigator at the Laboratory for Information and Decision Systems (LIDS). “When I realized that I could apply computer science, particularly AI and machine learning (ML), to the field of health, it felt like the perfect intersection of my interests.”
Currently, Ghassemi and her Healthy ML research group at LIDS delve deeply into enhancing the robustness of machine learning algorithms to ultimately promote safety and equity in health care.
Raised in Texas and New Mexico within an engineering-focused Iranian-American family, Ghassemi had ample inspiration to pursue a STEM career. While she reveled in puzzle-based video games—“The thrill of solving puzzles to unlock new levels was an enticing challenge”—her mother also nurtured her early engagement with advanced mathematics, encouraging her to perceive math as more than just basic arithmetic.
“While foundational skills like adding and multiplying are vital, an overemphasis can mask the fact that much of higher-level mathematics and science revolves around logic and problem-solving,” Ghassemi notes. “Thanks to my mother’s encouragement, I was aware of the exciting possibilities ahead.”
Beyond her mother, Ghassemi credits various mentors for their support throughout her educational journey. During her undergraduate studies at New Mexico State University, the director of the Honors College, Jason Ackelson—a former Marshall Scholar and now a senior advisor to the U.S. Department of Homeland Security—helped her secure a Marshall Scholarship. This led her to Oxford University, where she earned a master’s degree in 2011 and became intrigued by the innovative and fast-paced field of machine learning. Throughout her PhD at MIT, Ghassemi experienced valuable encouragement from both professors and peers, which fostered an environment of openness and acceptance that she strives to recreate for her own students.
Throughout her doctoral research, Ghassemi uncovered early evidence that biases in health data could obscure the effectiveness of machine learning models.
“I had been training models to predict outcomes from health data, and the prevailing thought was to utilize all available data. In image neural networks, we often found that the models could inherently learn relevant features, negating the need for manual feature engineering.”
During a committee meeting with Leo Celi, a principal research scientist at MIT’s Laboratory for Computational Physiology and a member of Ghassemi’s thesis committee, she was urged to assess the performance of her models across various patient demographics, including gender, insurance type, and self-reported race.
Upon investigation, Ghassemi discovered significant performance gaps. “Over the past decade, we’ve documented that these gaps pose substantial challenges to address, as they often reflect inherent biases in health data and established technical practices. Without careful consideration, models can perpetuate these biases,” she explains.
This realization sparked Ghassemi’s ongoing exploration into these critical issues.
Among her notable breakthroughs, she and her research team established that learning models could identify a patient’s race based on medical images, such as chest X-rays—an identification beyond the capability of human radiologists. They further discovered that models designed to optimize “average” performance did not perform well for women and minorities. This past summer, her team synthesized these insights to demonstrate that as a model improved its ability to predict a patient’s race or gender, the performance disparities for subgroups widened. Ghassemi’s research indicated that this bias could be mitigated by training models to recognize demographic differences rather than focusing solely on average performance. However, this adjustment must be applied at each unique site where a model is utilized.
“We emphasize that models trained for optimal performance in one hospital may not exhibit the same effectiveness in another setting. This distinction significantly impacts the applicability of these models in real-world health scenarios,” Ghassemi asserts. “While one facility may possess the resources to develop a well-performing model with fairness constraints, our findings highlight that such performance assurances do not transfer to new environments. A well-balanced model in one location may fail to operate effectively in another, which poses serious implications for the deployment of these technologies.”
Ghassemi’s work is deeply shaped by her identity.
“As a visibly Muslim woman and mother, these aspects of my identity influence my worldview, which in turn informs my research pursuits,” she shares. “My focus is on ensuring the robustness of machine learning models and understanding how their lack thereof intersects with existing biases. This connection is not coincidental.”
Ghassemi often finds inspiration outdoors—whether bike riding in New Mexico during her undergraduate days, rowing at Oxford, running as a PhD student at MIT, or now walking by the Cambridge Esplanade. She believes that breaking down complex problems into manageable parts can reveal assumptions that need revisiting to foster clearer understanding.
“In my experience, the most constraining factor in developing new solutions is the limits of your knowledge,” she states. “Sometimes, it’s challenging to see past your own understanding until you thoroughly investigate the components of a model or system and realize you might not fully grasp a critical element.”
While Ghassemi is undeniably passionate about her work, she makes a conscious effort to maintain perspective on life beyond academics.
“When you truly love your research, it’s easy for that passion to become intertwined with your identity—a challenge many academics face,” she reflects. “I strive to cultivate interests and insights that extend beyond my technical specialization.”
“Surrounding yourself with supportive individuals—friends, family, and colleagues—who encourage you to embrace a holistic identity is crucial!”
Having received numerous accolades and recognition for her pioneering efforts at the intersection of computer science and health, Ghassemi views life as a continuous journey.
“There’s a quote by the Persian poet Rumi: ‘You are what you are looking for.’ At every juncture in life, one must reinvest in self-discovery and actively guide that quest towards who you aspire to be,” she concludes.
Photo credit & article inspired by: Massachusetts Institute of Technology
