One of the first things I noticed while working with oncology data at Mayo Clinic was how difficult it can be to answer what seems like a simple question: Where did a patient’s cancer spread?
For many breast cancer patients, the answer is documented somewhere in the medical record. The challenge is finding it.
Why recurrence sites are difficult to identify
A radiology report may mention a suspicious lesion in the liver. A pathology report written months later may confirm metastatic disease. An oncologist’s note might discuss cancer involving the bone. Each document contains part of the story, but the information is often scattered across hundreds or even thousands of pages of clinical records.
For cancer registries and outcomes researchers, identifying recurrence sites is extremely important. Knowing whether cancer has spread to the bone, liver, lung, brain or other organs helps researchers study disease progression, evaluate treatments and understand long-term outcomes.
Yet much of this information still requires manual chart review. Trained reviewers spend hours reading through records and connecting clues from different parts of a patient’s history. It is meticulous work, but it is also time-consuming and difficult to scale.
As someone who develops artificial intelligence systems for health care, I wondered whether large language models could help.
Can AI understand clinical language?
Large language models have demonstrated impressive abilities at reading and generating text. But medical records present challenges that are very different from everyday language.
Clinical notes are filled with abbreviations, incomplete sentences and specialized terminology. Physicians often describe findings with varying levels of certainty. A statement such as “no evidence of metastatic disease” carries a completely different meaning from “findings concerning for metastatic disease.”
More importantly, understanding recurrence often requires connecting information across multiple documents written by different clinicians at different times.
In our study, my colleagues and I developed an AI framework designed to identify distant recurrence sites from unstructured clinical records of breast cancer patients. Rather than relying on structured database fields, the system reads the same types of documents that human reviewers use when determining where recurrence occurred.
The research is published in the Journal of Biomedical Informatics.
Testing beyond a single hospital
One lesson that researchers learn quickly in clinical AI is that success at one institution does not guarantee success elsewhere.
Medical centers differ in how they document care. Terminology varies. Clinical workflows vary. Even note-writing styles vary.
Because of this, we felt it was important to evaluate our approach beyond the environment where it was developed.
After developing the framework using data from Mayo Clinic, we performed external validation using data from Stanford Medicine. Seeing the system perform well across institutions was particularly encouraging because it suggested that the model was learning meaningful clinical patterns rather than simply adapting to local documentation habits.
Another finding surprised me.
Many people assume that larger AI models automatically perform better. In our study, however, a specialized framework designed specifically for recurrence-site identification outperformed several larger general-purpose language models. The result reinforced an idea that I believe is becoming increasingly important in health care AI: bigger models are not always better models for specialized clinical tasks.
Beyond breast cancer
We were also interested in whether the approach could generalize beyond breast cancer.
To explore this question, we applied the framework to prostate cancer data. Although the system had been developed for breast cancer recurrence, it adapted well to the new setting.
That finding suggested that the model was learning patterns associated with how recurrence is described in clinical language rather than learning only disease-specific terminology. While additional validation is still needed, it raises the possibility that similar approaches could eventually support multiple cancer types.
Why this matters
For me, the most exciting aspect of this work is not the AI model itself. It is the possibility of reducing the burden of manual chart review.
Hospitals generate enormous amounts of unstructured clinical text every day. Hidden within those records are answers to important research questions, but extracting those answers often requires substantial human effort.
If AI systems can reliably identify recurrence sites, researchers could spend less time searching through records and more time studying the factors that influence patient outcomes. Cancer registries could potentially process information more efficiently. Health systems could gain access to insights that are currently difficult and expensive to obtain.
Determining the best treatment for a patient will always require physician judgment. But there are many tasks that involve locating, organizing and summarizing information. These are areas where AI can serve as a valuable assistant.
Looking ahead
Every day, hospitals generate clinical narratives that capture valuable knowledge about disease progression, treatment response and patient outcomes. Much of that information remains difficult to analyze at scale.
The challenge now is moving beyond proof-of-concept studies and developing tools that clinicians, researchers and health systems can trust and use in practice. That will require rigorous validation, careful evaluation across diverse populations and close collaboration between AI researchers and medical experts.
The technology is advancing rapidly. What excites me most is the possibility that these tools can help transform information that is currently hidden in medical records into knowledge that improves cancer research and, ultimately, patient care.