Adam Yala, PhD, assistant professor of Computational Precision Health and EECS at UC Berkeley and University of California, San Francisco (UCSF), delivered a captivating talk on the transformative potential of machine learning in cancer care at the SRG Pillar meeting for the Center for Intelligent Imaging (ci2). Dr. Yala shared his research on developing computational methods for predicting cancer risk and creating adaptive screening policies. His approach emphasizes precision and practicality, bridging academic innovation and clinical application.
Dr. Yala began by framing cancer care as a computational problem, "Given all we understand about a patient, which is many gigabytes of data...longitudinal imaging, pathology, and electronic health records—we aim to figure out who to intervene on, how, and when to maximize outcomes." This approach critiques the one-size-fits-all paradigm in cancer screening, such as annual mammograms starting at age 40, which often results in what Dr. Yala described as "dual harms"—missing advanced cancers in some patients while subjecting others to unnecessary procedures and anxiety.
Dr. Yala highlighted the advancements of Mirai, his breast cancer risk prediction tool, which uses image-based models rather than relying solely on traditional clinical variables like age and family history. "Mirai has already shown significant improvements, outperforming clinical models in risk prediction," Dr. Yala noted. Validated in 43 hospitals across 14 countries, the model has also been incorporated into multiple clinical trials, including one starting at San Francisco General Hospital. "These trials aim to allocate interventions, such as MRIs, more effectively," Dr. Yala explained, emphasizing the importance of real-world testing.
While these strides are remarkable, Dr. Yala stressed that current methods only scratch the surface of what's possible. "We've moved from bits to megabytes, but the real-world data is in the gigabyte to terabyte range. Modeling everything requires entirely new approaches."
One of the key challenges Dr. Yala's team addressed is structuring the immense volume of patient data to make it useful for machine learning models. "To understand a patient's cancer journey, we need to extract critical information from tens of thousands of pathology and radiology reports," he shared. To achieve this, his team developed Strata, a low-cost, low-code tool for extracting clinical data. Remarkably, it achieves human-level accuracy with minimal annotations. "With Strata, we're able to curate datasets at scale, which then feed into our models," Dr. Yala explained.
Scaling these models further, Yala introduced Tower, a novel neural network architecture designed to handle massive datasets efficiently. Unlike traditional models that require prohibitive computational resources, Tower reduces costs significantly. "With Tower, we can run million-token experiments on a single machine—something that was unthinkable before," Dr. Yala noted. This efficiency enables the exploration of longitudinal imaging and other rich datasets, setting the stage for more comprehensive cancer risk predictions.
Another forward-thinking aspect of this research is reimagining clinical trials to keep pace with rapidly improving machine learning models. "By the time a trial finishes, the model it tested is often outdated," Dr. Yala shared. To address this, his team proposed adaptive trial frameworks that allow newer model iterations to build upon prior results. This strategy could save substantial resources and accelerate the deployment of better tools in clinical settings.
Dr. Yala concluded with a vision for integrating diverse data sources—imaging, pathology, and electronic health records—to tackle complex questions in cancer care. "If we do this well, we can not only design better screening guidelines but also identify patients who won't respond to certain therapies," he said. His team's ultimate goal is to expand these methods across multiple cancer types, including lung, pancreatic, and prostate cancers.
Through his methodical and scalable approach, Adam Yala is driving cancer screening and care toward greater precision, efficiency, and accessibility. His research exemplifies the promise of computational precision health in reshaping clinical outcomes.
To learn more about the upcoming SRG Pillar meetings, visit the ci2 events page.
