Published May 30, 2026 | Version 1.0.1

LLMSmokeDetector: a two-stage open-source LLM pipeline for tobacco smoking status extraction from clinical discharge summaries

  • 1. Institute for Informatics, Data Science & Biostatistics, Washington University School of Medicine in St. Louis

Contributors

Data curator:

  • 1. Washington University School of Medicine in St. Louis

Description

Source code for a two-stage large language model (LLM) inference pipeline that extracts tobacco smoking status (Smoker, Never smoker, Unknown) from free-text hospital discharge summaries. Stage 1 generates an unstructured classification per note; Stage 2 maps that output to a single structured label using constrained decoding. Evaluation computes multiclass and per-class metrics with bootstrap confidence intervals and confusion matrices. Companion code for the JAMIA Open Brief Communication comparing six lightweight open-source LLMs (Llama-3 1B-70B, gpt-oss-20B, MedGemma-27B) against the cTAKES NLP system.

Files

ddav118/LLMSmokeDetector-v1.0.1.zip

Files (29.9 kB)

Name Size
md5:65757fe3f5626028923d0f0c29404ccb
29.9 kB Preview Download

Additional details

Related works

Is supplement to
Journal article: 10.1093/jamiaopen/ooaf182 (DOI)

Software

Repository URL
https://github.com/ddav118/LLMSmokeDetector
Programming language
Python