Unstructured Data Are Superior to Structured Data for Eliciting Quantitative Smoking History From the Electronic...

Area of Science:

• Medical informatics
• Computational linguistics
• Public health

Background:

• Accurate identification of smoking status is crucial for lung cancer screening eligibility.
• Traditional structured data often lacks detailed smoking history, hindering cohort identification for low-dose computed tomography (LDCT).
• Natural language processing (NLP) offers a potential solution for extracting nuanced clinical information from unstructured text.

Purpose of the Study:

• To develop and validate a method for extracting smoking status and quantitative smoking history from clinician notes.
• To facilitate the identification of patient cohorts eligible for LDCT screening for early lung cancer detection.
• To improve the accuracy and efficiency of determining LDCT eligibility based on smoking criteria.

Main Methods:

• Utilized natural language processing (NLP) and named entity recognition on clinician notes from the MIMIC-III database.
• Developed algorithms to extract quantitative smoking history, including pack-years and quit dates.
• Validated extracted data against manual chart reviews, calculating an F-score for accuracy.

Main Results:

• NLP identified 1,930 ever-smokers from clinician notes, a significant increase compared to structured data.
• The method successfully identified 276 patients eligible for LDCT screening based on smoking and age criteria, using USPSTF guidelines.
• The F-score for identifying eligible patients using this NLP-based approach was 0.88, demonstrating high accuracy.

Conclusions:

• Unstructured clinical data, when processed by NLP, can accurately identify specific patient cohorts for LDCT screening.
• This NLP-driven method enhances the precision of cohort identification for lung cancer early detection programs.
• The findings support the integration of NLP tools in electronic health records to optimize screening eligibility determination.