An Order-Sensitive Hierarchical Neural Model for Early Lung Cancer Detection Using Dutch Primary Care Notes and...

Area of Science:

• Medical Informatics
• Oncology
• Natural Language Processing

Background:

• Timely lung cancer detection is crucial for improving patient outcomes.
• Primary care settings offer opportunities for early diagnosis through analysis of clinical notes.
• Leveraging unstructured text data can enhance prediction models.

Purpose of the Study:

• To develop and validate prediction models for early lung cancer detection using free-text consultation notes.
• To assess the utility of models incorporating text alone versus text combined with clinical variables.
• To evaluate model performance in terms of discrimination, positive predictive value (PPV), and calibration.

Main Methods:

• Utilized a large dataset from general practices (2002-2021) including patients over 30 with free-text notes.
• Developed hierarchical models using attention and bidirectional long short-term memory networks.
• Trained and tested models on separate datasets, excluding data close to diagnosis, and incorporating clinical variables.

Main Results:

• The combined model (text + clinical variables) achieved an AUROC of 0.91 and AUPRC of 0.05 on the test set.
• Positive predictive value (PPV) was 0.034, with good calibration observed.
• Identifying one lung cancer patient required additional testing for 29 high-risk individuals.

Conclusions:

• The developed models demonstrate excellent discrimination for early lung cancer detection by analyzing text structure.
• Combining clinical variables with text slightly enhanced model performance.
• The models show promise for clinical practice, warranting further external validation and implementation studies.