VTE-LLM

This repository contains source code for our manuscript: "Large language model (LLM) based natural language processing (NLP) improves detection of overall and recurrent venous thromboembolism (VTE) in patients with cancer"

Requirements

• Python 3.9
• For LLM model, we used Claude Sonnet 4.5 through AWS Bedrock api.
• To run the pipeline, download the VTE-BERT models at this link, and install Transformers 4.55.0

Preprocessing:

To reduce the number of input tokens, we used this preprocessing:

For more detail visit this repo: https://github.com/Ang-Li-Lab/NLPMed-Engine

Methods

• VTE-LLM process:

  

• VTE-BERT + VTE-LLM process: VTE-BERT plays as a screening tool that filters out patients without evidence of VTE at the note level, reducing the number of patients requiring VTE-LLM inference.

  

Results

VTE-LLM on distinct level results at Harris Health System:

Classified \ Actual	Positive	Negative	Metric
Positive	245 (TP)	14 (FP)	PPV = 95%
Negative	16 (FN)	192 (TN)	NPV = 92%
Metric	Sensitivity = 94%	Specificity = 93%	467 events

VTE-BERT+VTE-LLM on distinct level results at Harris Health System:

Classified \ Actual	Positive	Negative	Metric
Positive	239 (TP)	12 (FP)	PPV = 95%
Negative	22 (FN)	199 (TN)	NPV = 90%
Metric	Sensitivity = 92%	Specificity = 94%	472 events

For more results including external validations, incident and recurrent VTE detection metrics, please refer to our full paper.

Examples:

We provided some examples how to run the pipeline, the direct LLM, and post processing in notebooks/example_pipeline.ipynb

Citation

Please cite our work when using the code, methodology, or results in publications, presentations:

@article{pham2026vte,
  title={Large language model (LLM) based natural language processing (NLP) improves detection of overall and recurrent venous thromboembolism (VTE) in patients with cancer},
  author={Dang Pham*, Aidan Boyne*, Jennifer La, D. Siamack Ayandeh, Barbara Lam, Mrinal Ranjan, Juan Bueno, Emma Kitchel, Melkon Harutunian, Gongliang Zhang, Nathanael Fillmore, Omid Jafari, Ang Li},
  journal={...},
  publisher={...},
  year={...},
  doi={...}
}