Patient representation learning and interpretable evaluation using clinical notes.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Processing

Background:

• Clinical notes contain valuable patient information but are often unstructured.
• Learning effective patient representations is crucial for downstream clinical tasks.
• Current methods may not fully capture the complexity of clinical text.

Purpose of the Study:

• To develop and evaluate task-independent dense patient representations from clinical notes.
• To assess the transferability of these representations across different predictive tasks.
• To enhance model interpretability for clinical data analysis.

Main Methods:

• Utilized stacked denoising autoencoders and paragraph vector models for dense representation learning.
• Compared dense representations against sparse bag-of-words models.
• Evaluated performance on predicting patient mortality, diagnostic category, and gender.
• Investigated the impact of medical concept identification versus bag-of-words features.
• Developed techniques for model interpretability, including feature encoding and sensitivity analysis.

Main Results:

• Learned generalized dense representations significantly outperformed sparse representations, especially with limited positive instances and weak lexical features.
• Concept identification did not improve classification performance compared to bag-of-words.
• Novel interpretability techniques successfully identified the most influential features for classification tasks.

Conclusions:

• Task-independent dense patient representations derived from clinical notes offer superior performance over traditional sparse methods.
• The transferability of these representations enhances their utility across various clinical prediction tasks.
• Proposed interpretability methods provide valuable insights into the factors driving model predictions.