By Sujatha Sagiraju, Senior Vice President of Engineering, Clinical Intelligence, and AI Evaluation, and Swapna Abhyankar, MD, physician informaticist at Truveta
Today, we’re excited to share that our latest paper on extracting cancer stage from clinical notes has been published in JCO Clinical Cancer Informatics (JCO CCI). The paper describes how we developed AI models to extract cancer stage data from unstructured clinical notes.
The results presented in the paper are important, but they are only part of the story. This work also established a foundation for extracting more of the rich clinical context oncology researchers need, across large and diverse populations.
Importance
Buried within unstructured narratives is some of the richest clinical information that exists in oncology. Details like cancer stage, performance status, histological grade and type of cancer, tumor markers, metastasis, medications, and disease progression are critical for oncology research. We started with staging data for five cancer types (as reported in our paper), and have since expanded those methods to extract all the above variables and more for all solid tumors from clinical documentation. Extracting and normalizing these variables and combining them with structured electronic health record (EHR) data has the promise to significantly accelerate cancer research.
For patients and families of patients living with cancer, progress cannot come fast enough. We do this work for them in the spirit of advancing our mission of Saving Lives with Data.
Challenge
Clinical documentation is nuanced, inconsistent, and highly contextual. Different types of notes have variable content that is documented in a variety of formats, with abbreviations, local naming conventions, typos, and data that are copied and pasted from previously written notes. Data may be incomplete, and any available data must be extracted along with the context needed to interpret them, such as the verification status of a diagnosis, whether confirmed or refuted, and its temporality.
There is no established playbook for extracting complex variables from billions of clinical notes. Success required solving three challenges simultaneously:
1.Understanding the clinical complexity
2. Developing AI models that could reason over real-world documentation
3. Building engineering systems that could process notes at this scale while producing data that researchers and clinicians can trust
Scale
We deliberately started with a defined problem: extracting cancer stage for five cancer types from more than two million notes representing approximately 200,000 patients. Starting with a focused set of cancers allowed us to evaluate model performance, identify documentation patterns, refine annotation guidance, and improve our methods before expanding further. In the last few months, we have expanded staging extraction to all solid tumors across more than 750,000 patients. And we are still just beginning.
Team
This work was only possible because of close collaboration across clinical, AI, and engineering teams. Clinicians helped AI scientists understand the subtle ways cancer characteristics are documented in practice. AI scientists translated those clinical insights into robust extraction models by fine-tuning large language models. Engineers built the infrastructure needed to process billions of notes reliably and at scale.
No single discipline could have solved this problem independently. Clinical expertise was necessary to define what the information meant, AI was needed to identify it consistently, and engineering was needed to make the approach usable in the real world.
We’re incredibly proud of what this multidisciplinary team has accomplished, and even more excited about how this work can help contribute to our mission of Saving Lives with Data.
