Comparing deep learning and concept extraction based methods for patient phenotyping from clinical narratives

Area of Science:

• Medical informatics
• Natural Language Processing
• Machine Learning

Background:

• Accurate patient cohort identification from electronic health records is vital for research.
• Clinical narratives contain crucial, yet often unstructured, patient data.
• Traditional methods rely on clinician-defined concepts and machine learning.

Purpose of the Study:

• To compare Convolutional Neural Networks (CNNs) against concept extraction methods for patient phenotyping.
• To evaluate the performance of deep learning models in identifying medical conditions from clinical text.
• To assess the interpretability of CNNs for clinical applications.

Main Methods:

• Secondary analysis of 1,610 discharge summaries from the MIMIC-III database.
• Comparison of CNN-based text classification with concept extraction methods.
• Evaluation of model performance using F1-score and Area Under the ROC Curve (AUC).
• Assessment of model interpretability through salient phrase extraction.

Main Results:

• CNNs outperformed concept extraction methods in nearly all ten phenotyping tasks.
• Improvements of up to 26 in F1-score and 7 percentage points in AUC were observed with CNNs.
• CNNs demonstrated comparable or superior interpretability by highlighting relevant phrases.

Conclusions:

• CNNs represent a powerful alternative for patient phenotyping and cohort identification.
• Deep learning models can effectively leverage rich language representations in clinical narratives.
• The approach can aid clinicians in chart review and billing code extraction.