Interpretable Probabilistic Latent Variable Models for Automatic Annotation of Clinical Text.

Area of Science:

• Computational linguistics
• Medical informatics
• Machine learning

Background:

• Automatic annotation of clinical text is crucial for data analysis.
• Existing probabilistic models may lack interpretability or optimal performance.

Purpose of the Study:

• To propose novel interpretable probabilistic latent variable models for clinical text annotation.
• To enhance the accuracy and interpretability of automatic text annotation.

Main Methods:

• Latent Class Allocation (LCA) model: learns class-specific term distributions.
• Discriminative Labeled Latent Dirichlet Allocation (DL-LDA) model: decomposes class distributions into topics.
• Comparison with Naïve Bayes and Labeled LDA.

Main Results:

• Proposed LCA and DL-LDA models outperform Naïve Bayes and Labeled LDA.
• The models demonstrate superior performance in annotating motivational interview transcripts.
• The output of LCA and DL-LDA is interpretable by clinical practitioners.

Conclusions:

• LCA and DL-LDA offer a significant advancement in interpretable automatic clinical text annotation.
• These models provide accurate and understandable annotations for clinical text data.
• The developed models are valuable tools for clinical text analysis and interpretation.