SmokeBERT: A BERT-based Model for Quantitative Smoking History Extraction from Clinical Narratives to Improve Lung...

Area of Science:

• Computational linguistics
• Medical informatics
• Public health

Background:

• Tobacco use is a major risk factor for cancer and cardiovascular diseases.
• Electronic health records often lack detailed quantitative smoking data (e.g., pack years) in structured fields.
• Accurate smoking history is vital for disease risk assessment and lung cancer screening (LCS) eligibility.

Purpose of the Study:

• To develop and evaluate SmokeBERT, a BERT-based natural language processing (NLP) model for extracting detailed quantitative smoking histories from clinical narratives.
• To improve the accuracy of identifying patients eligible for lung cancer screening.

Main Methods:

• Fine-tuning a BERT-based model (SmokeBERT) on clinical notes to extract granular smoking data.
• Comparing SmokeBERT's performance against a state-of-the-art rule-based NLP model.
• Evaluating extraction accuracy using F1 scores and identification of LCS-eligible patients.

Main Results:

• SmokeBERT achieved a superior F1 score (0.97) compared to the rule-based model (0.88) on a hold-out test set.
• SmokeBERT significantly improved the identification of LCS-eligible patients (e.g., 98% vs. 60% for ≥20 pack years).

Conclusions:

• SmokeBERT demonstrates superior performance in extracting detailed smoking histories from clinical text.
• This NLP advancement can enhance the accuracy of lung cancer screening eligibility determination.
• Future work aims to expand SmokeBERT's capabilities to multilingual and larger-scale applications.