SmokeBERT: A BERT-based Model for Quantitative Smoking History Extraction from Clinical Narratives to Improve Lung Cancer Screening
Xue, Y.; Zhu, Y.; Zhuang, L.; Oh, Y.; Taira, R.; Aberle, D. R.; Prosper, A. E.; Hsu, W.; Lin, Y.
Show abstract
Tobacco use is a critical risk factor for diseases such as cancer and cardiovascular disorders. While electronic health records can capture categorical smoking statuses accurately, granular quantitative details, such as pack years and years since quitting, are often embedded in clinical narratives. This information is crucial for assessing disease risk and determining eligibility for lung cancer screening (LCS). Existing natural language processing (NLP) tools excelled at identifying smoking statuses but struggled with extracting detailed quantitative data. To address this, we developed SmokeBERT, a fine-tuned BERT-based model optimized for extracting detailed smoking histories. Evaluations against a state-of-the-art rule-based NLP model demonstrated its superior performance on F1 scores (0.97 vs. 0.88 on the hold-out test set) and identification of LCS-eligible patients (e.g., 98% vs. 60% for [≥]20 pack years). Future work includes creating a multilingual, language-agnostic version of SmokeBERT by incorporating datasets in multiple languages, exploring ensemble methods, and testing on larger datasets.
Matching journals
The top 4 journals account for 50% of the predicted probability mass.
Similar papers in this journal
Similar papers in this journal
Similar papers in this journal
- A Deep Learning Method to Detect Opioid Prescription and Opioid Use Disorder from Electronic Health Records 94%
- Image and structured data analysis for prognostication of health outcomes in patients presenting to the Emergency Department during the COVID-19 pandemic 90%
- Assessing the effects of data drift on the performance of machine learning models used in clinical sepsis prediction 90%
"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.