Authors: Swapna Abhyankay, MD ⊕Truveta, Inc, Bellevue, WA, Rajesh M Rao, MS ⊕Truveta, Inc, Bellevue, WA, Mehraveh Salehi, PhD ⊕Truveta, Inc, Bellevue, WA, Jennifer J Liang, MD ⊕Truveta, Inc, Bellevue, WA, Jamalynne Deckard, MS ⊕Truveta, Inc, Bellevue, WA, Sujatha Sagiraju, MS, MBA ⊕Truveta, Inc, Bellevue, WA
- TLM-Oncology extracted 2.7 million detailed cancer staging records from more than 2.1 million clinical notes for 217,768 patients
- The model captured overall stage, TNM values, stage type, staging method, diagnosis, and timing across five cancer types
- Precision remained high for breast and prostate cancers, even though they were not included in oncology-specific training
- By converting staging details from free-text notes into structured records, TLM-Oncology could make large-scale real-world oncology research more feasible
Cancer staging is essential for oncology research, but it is often documented in free-text clinical notes rather than structured fields in the electronic health record (EHR). This makes it difficult to study cancer populations at scale, especially when researchers need detailed stage, tumor, node, metastasis, stage type, staging method, and timing information.
In a new study published in JCO Clinical Cancer Informatics, the Truveta team developed and evaluated Truveta Language Model for Oncology (TLM-Oncology), a large language model (LLM) designed to extract detailed cancer staging information from clinical documentation.
Building TLM-Oncology for detailed staging extraction
Researchers developed and evaluated TLM-Oncology using de-identified clinical notes from patients with one of five cancer types: bladder, breast, cervical, colorectal, and prostate.
Clinical terminologists created a reference dataset of 700 notes. TLM-Oncology was trained and validated using notes from patients with bladder, cervical, and colorectal cancers, then tested on those cancers and two cancers it had not encountered during oncology-specific training: breast and prostate. The model extracted:
- Overall cancer stage
- Tumor, node, and metastasis values, known as TNM
- Stage type and staging method
- The associated cancer diagnosis
- The timeframe when staging was assessed
TLM-Oncology achieved high precision across cancer types
Across the full study population, TLM-Oncology extracted 2,747,800 staging records for 217,768 patients from 2,149,136 notes.
The model achieved high precision across cancer types and note types, including pathology reports, clinical notes, and diagnostic imaging. At the relational level—where every attribute and relationship in a staging record had to be correct—precision ranged from:
- 77 to 1.00 for bladder, cervical, and colorectal cancers
- 83 to 1.00 for breast and prostate cancers, which were not included in oncology-specific training
Researchers found no statistically significant difference in F1 scores between the cancers used for training and the two previously unseen cancers. This finding suggests that the model may transfer to additional cancer types, although broader validation is still needed.
Performance varied by cancer, note type, and staging attribute. The model generally performed better on more consistently structured pathology reports than on clinical notes and imaging reports. Most errors reflected missed information rather than incorrect extractions.
Why it matters
Cancer staging helps researchers define disease severity, compare treatment patterns, and compare outcomes. But when staging is available only in notes, it can be difficult or impossible to use at scale. By converting data from unstructured notes into structured records, TLM-Oncology could help researchers access critical cancer variables for real-world oncology research. The model also captures more than an overall stage, preserving details such as TNM values, stage type, diagnosis, and timing that can change how a staging record should be interpreted.
Read the full study
Large Language Models to Extract Cancer Staging Data From Clinical Documentation at Scale, JCO Clinical Cancer Informatics
