Regina Barzilay, PhD, worked in the area of artificial intelligence (AI) from the beginning of her career in computer science she has applied it to translations of dead languages—a natural fit for someone with a master’s degree in natural language processing—like Linear B, and optimizing strategy in video games like Civilization II. In 2014, however, this year’s recipient of the Wallace H. Coulter Lectureship Award had an experience that led her to turn toward medicine: she was diagnosed with breast cancer. In her plenary session yesterday afternoon, “Artificial Intelligence in the Clinic,” she described how her experiences as a patient led her to recognize the trove of data captured, but not fully utilized, and how she is using AI to change that.

With little illness prior to her diagnosis, this was Barzilay’s first major encounter with the healthcare system. She was struck by the primitive use of information. Everything relied upon clinical trial data, but what about the experiences of every patient treated in accordance with a trial’s findings? These were not being captured. Could AI find the data locked within clinic notes? Could machine learning, a type of AI, not only be trained to identify cancer in mammograms, but also to identify patients, who did not have breast cancer today but would in a few years?

Barzilay, distinguished professor in the department of electrical engineering and computer science at the Massachusetts Institute of Technology (MIT), described machine learning as “enabling a machine to map the input conditions to the output conditions.” Algorithms are trained to determine the pathway between points A and B rather than being instructed explicitly. Barzilay described how, as her journey led her into the clinic, she leveraged her experience in natural language processing to train an algorithm to extract relevant tumor characteristics from three decades’ worth of pathologists’ notes on breast cancer patients in the Massachusetts General Hospital (MGH) system. Such annotation allowed rapid identification of historical cases of interest.

Trained AI models can detect malignancy on mammograms as well, or better, than trained human eyes. Working with colleagues at MIT and MGH, Barzilay next developed an AI system, named Mirai, to extract information from mammograms and other risk factors using large datasets from locations around the world, enabling it to determine a patient’s risk of developing breast cancer within the next five years. Mirai has been part of decision support at MGH since 2018, and was invaluable during the COVID-19 pandemic, when screening centers in Boston were closed or operating at reduced capacity. She helped identify patients who should not postpone their screenings.

Barzilay is adapting the risk assessment and diagnostics tools developed for breast cancer for lung and prostate cancers, but she also described how she continues to apply AI to gaps in the medical landscape. Novel antimicrobials are needed; Barzilay applied machine learning to identify a novel antimicrobial, halicin, which also has a novel mechanism of action. Recently, she leveraged AI to identify two drugs with synergistic activity against SARS-CoV-2 in vitro.

As she looks toward the future, Barzilay expressed excitement for the prospects of personalized medicine. To her, personalized medicine extends beyond the impact of genetics, where discussions of personalized medicine often end. Laboratory data, analyzed by artificial intelligence—possibly, but not necessarily, alongside genetic information—offers the possibility of offering sound decision support for clinicians. As AI rises in profile in the clinic, new opportunities for AI-based support may reveal themselves, and Barzilay, her colleagues, and those inspired by them will help us make the most of those opportunities.