Back

From text to tables: Zero-shot extraction of structured clinical data from free-text CT scan reports using foundational large language models

Hongslo, A.; Gupta, A.; Nguyen, Q.; Caldwell, J.; Choi, B.; Harvey, C. J.; Thompson, J.; Mazzotti, D.; Yao, Z.; Noheria, A.

2025-10-07 health informatics
10.1101/2025.10.05.25337366 medRxiv
Show abstract

BackgroundLarge language models (LLMs) are being explored for multiple applications in medical research, including medical text classification. We evaluate the performance of 5 off-the-shelf LLMs for classifying free-text CT angiography reports for pulmonary embolism (PE)- related diagnostic labels. MethodsWe assessed 1,025 manually labeled CT reports using 5 LLMs (ChatGPT-4o, Llama 3.3 70b, Llama 3.1 8b, Llama 3.2 3b, Mixtral 8x7b) with zero-shot prefix prompts. Labels included acute PE, bilateral PE, and large PE. Voting ensemble models combining multiple LLM outputs were also tested. ResultsLlama 3.3 70b and ChatGPT-4o outperformed smaller models for all classification tasks. Highest accuracies were 96.6% (acute PE), 92.7% (bilateral PE), and 82.6% (large PE). Voting ensemble models offered no or minimal improvement in classification performance. ConclusionsOff-the-shelf LLMs, particularly larger ones, can classify free-text reports with high accuracy using simple prompts. Further work is needed to optimize prompting strategies and evaluate hybrid approaches.

Matching journals

The top 3 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.