Using AI Chains for SNOMED CT entity extraction from free text clinical notes

LLM Chains allow us to divide a complex task into a Chain of smaller tasks, where the output of one prompt becomes the input of the next.

This is not significantly slower, despite requiring a greater number of calls/requests to the LLM, because short outputs take a proportionally shorter time to generate. Although the input prompt tokens also need to be processed each time, they are less time-expensive than output generated tokens in the case of LLaMA 2.

Chain design

Note: the diagram below is currently outdated and will be refreshed in a future update.

How to run the Entity Extractor

The entity_extractor.py script extracts clinical entities from free text using LLMs and maps them to SNOMED CT. It supports openai, llama, and bard as API backends.

For OpenAI, the project is configured for the current Python SDK and uses the modern Responses API (with a compatibility fallback), including support for ChatGPT-5 family models (for example gpt-5-mini).

Prerequisites

To run entity_extractor.py, you will need:

• Python 3.x
• OpenAI: an API key for OpenAI API (see https://platform.openai.com/)
• Llama-2 (optional): llama.cpp installed, and a local copy of Llama-2 quantized for use in llama.cpp (see https://github.com/ggerganov/llama.cpp)

Usage

Follow these steps:

• Clone the repository to your local machine.
• Open the project in your preferred code editor.
• Install dependencies: pip install -r requirements.txt
• If you need to use OpenAI models, create a local openai.key file in the project root and paste your API key there.
• Run the script with:

python entity_extractor.py --api <api> --model <model> --sentences <path>

CLI parameters

• -a, --api: backend to use (openai, llama, bard). Default: openai.
• --model: model identifier.
  • OpenAI example: gpt-5-mini
  • Llama example: local model file path
• --sentences: path to a text file containing one clinical sentence per line.
• --n: optional limit for number of non-empty, non-comment lines to process.
• --mode: output mode.
  • verbose (default): shows matching stages and detailed output.
  • summary: shows only Processing sentence X..., then final colored sentence output and per-line score.

Optional environment variables

• OPENAI_DEBUG_CACHE=1: enables request/response cache debug logs from the OpenAI client layer.

Examples:

python3.10 entity_extractor.py --api llama --model /my-drive/models/llama2/llama-2-13b-chat/ggml-model-q4_0.bin --sentences=example_cases/clinical_text.txt

python3 entity_extractor.py --api openai --model gpt-5-mini --sentences=example_cases/clinical_text.txt

OPENAI_DEBUG_CACHE=1 python3 entity_extractor.py --api openai --model gpt-5-mini --sentences=example_cases/clinical_text.txt --mode summary

References

• Wu, T., Terry, M., & Cai, C. J. (2022, April). AI chains: Transparent and controllable human-ai interaction by chaining large language model prompts. In Proceedings of the 2022 CHI conference on human factors in computing systems (pp. 1-22). https://doi.org/10.1145/3491102.3517582
• Cursor https://www.cursor.so/blog/llama-inference#user-content-fn-llama-paper [Retrieved 08/2023]
• LLaMA 2 – Meta AI https://ai.meta.com/llama/ [Retrieved 08/2023]