SmokeBERT: A Bidirectional Encoder Representations From Transformers-Based Model for Quantitative Smoking History...

Area of Science:

• Natural Language Processing (NLP)
• Clinical Informatics
• Biomedical Data Extraction

Background:

• Tobacco use is a primary risk factor for cancers and other diseases.
• Accurate quantitative smoking data (pack-years, quitting time) is crucial for risk assessment and lung cancer screening (LCS).
• Existing NLP tools face challenges in extracting granular smoking details from clinical narratives.

Purpose of the Study:

• To develop and evaluate a novel NLP model, SmokeBERT, for precise extraction of quantitative smoking history from clinical texts.
• To improve the accuracy of identifying patients eligible for lung cancer screening based on smoking history.

Main Methods:

• Four Bidirectional Encoder Representations from Transformers (BERT)-based models (BERT, BioBERT, ClinicalBERT, MedBERT) were fine-tuned on clinical narratives.
• Models were trained on 3,261 sentences mentioning smoking history to extract six quantitative variables.
• ClinicalBERT was selected, fine-tuned into SmokeBERT, and validated on holdout and external datasets.

Main Results:

• SmokeBERT achieved superior performance in smoking data extraction compared to state-of-the-art NLP models (F1 scores: 0.97 on holdout, 0.86 externally).
• SmokeBERT accurately identified patients eligible for lung cancer screening (e.g., ≥20 pack-years: 97%, ≤15 years since quitting: 100%).
• Performance surpassed existing rule-based and large transformer models.

Conclusions:

• SmokeBERT, a fine-tuned BERT model, effectively extracts detailed quantitative smoking histories from clinical data.
• The model demonstrates significant potential for enhancing clinical risk assessment and patient screening.
• Future research will focus on larger datasets and multilingual, language-agnostic versions.