When I entered the world of clinical genomics, I expected long days parsing through journal articles, tidying spreadsheets of patient data and running Python scripts late into the night. I was prepared to code, not to listen.
That changed when I began working with clinical notes — messy, personal and inconsistent footprints left by physicians. Buried in them were powerful clues: a family history of breast cancer, a BRCA1 mutation, a hesitant mention of a diagnosis. These fragments shaped diagnosis, care and discovery.
The problem? These clues are easy to miss. Manually reviewing thousands of records is slow, subjective and error-prone. That’s where artificial intelligence, or AI, specifically natural language processing, or NLP, began to change my work.
As a bioinformatics graduate student at Michigan Medicine, I build hybrid NLP systems that bridge structure and storytelling — algorithms trained to “read” closely, knowing that “positive” doesn’t mean hopeful and that “no history” could be a red flag.
Our models combine classic machine learning methods, like conditional random fields, with domain-specific rules and genetic lexicons. The rules, based on medical knowledge, prevent errors like flagging “no family history of breast cancer” as a risk. The lexicons act as specialized dictionaries of gene names, mutations and terms that help the system extract key variants, inheritance patterns and clinical details from messy notes.
The result: faster, more consistent identification of patients who may benefit from genetic testing or clinical trial enrollment. For example, if a note states, “Mother diagnosed with breast cancer at 42; patient reports positive BRCA1 mutation,” our system flags this as a high-risk case and alerts clinicians to recommend genetic counselling or appropriate clinical trials. This cuts delays, reduces subjectivity and helps ensure fewer patients slip through the cracks.
What surprised me most wasn’t the power of these tools, but how they’ve challenged my assumptions. What does “accuracy” mean in medicine? What errors are acceptable? In clinical NLP, a mislabel isn’t minor — it means someone gets missed. If the system misreads “no history of colon cancer” as “history of colon cancer,” a patient might face unnecessary testing, or worse, real risks could be overlooked. In genetic medicine, such errors can delay diagnosis and treatment with serious consequences.
There’s still a long road ahead. I’m exploring how to move beyond extraction toward interpretation. Can we build models that not only find genetic risks but summarize them meaningfully? Could AI reveal disparities in access to genetic counseling or care?
AI hasn’t replaced my work — it’s reshaped it. It’s made me more careful, collaborative and curious. In a field where the data is deeply human, the greatest thing AI has taught me is how to listen.