Labeling Clinical Notes With Large Language Models: Retrospective Benchmarking Study
Related Experiment Video
Updated: Nov 12, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Pritam Mukherjee1, Michael Nazmifar1, Vivek Batheja1
1Imaging Biomarkers and Computer Aided Diagnosis Lab, Radiology and Imaging Sciences, National Institutes of Health Clinical Center, 10 Center Dr, Building 10, Rm 1C224, Bethesda, MD, 20892, United States, 1 301-402-5486.
Background:
Clinical notes are rich sources of clinical information and are routinely used in clinical practice to infer past diagnoses and other comorbidities. Large language models (LLMs) may simplify the task by extracting this information from clinical notes.
Objective:
The study aims to evaluate the performance of state-of-the-art LLMs on the task of labeling English clinical notes.
Methods:
This institutional review board-approved retrospective study comprised 380 clinical notes obtained from 2 institutions in the United States. The notes were manually labeled for the presence or absence of 29 conditions (simplified Elixhauser Comorbidity Index) by 2 human readers-a radiology resident and a medical student under the supervision of a senior radiologist. Forty-six LLMs, including 4 closed-source commercial models from the GPT and Claude families, and 42 open-weights LLMs from the Llama (13 models) and the Qwen (29 models) families were used to label each note. The GPT-o3-mini-high reasoning model was used to parse the free-text model responses to binary yes/no answers. Using the manual labels as the reference standard, the recall, precision, and F1-scores for the models were evaluated. The effect of factors such as the length of the notes and the family, size, and release recency of the LLMs was analyzed using generalized linear models.
Results:
The top-performing LLMs included the GPT and Claude models, and the larger open-weights models included Llama 3 70B, Llama 3.1 70B, Llama 3.1 405B, Llama 3.3 70B, Qwen 1.5 72B, Qwen 1.5 110B, Qwen 2 72B, and Qwen 2 VL 72B, which achieved F1-scores above 70%. The most recent and higher-capacity releases of the LLMs have been improving mainly in their precision. The evaluation of 29 conditions showed a large variance in performance depending on the condition.
Conclusions:
In this evaluation with 46 LLMs, larger open-weights LLMs such as the Llama 3.3 70B performed on par with or better than the GPT and Claude models, achieving F1-scores above 70%. The performance of open-weights models improved significantly with model size and recency. These results suggest that locally run open-weights LLMs can be potentially used to label clinical notes.