Toward a Neural Semantic Parsing System for EHR Question Answering.

Area of Science:

• Natural Language Processing
• Health Informatics
• Machine Learning

Background:

• Clinical semantic parsing (SP) is crucial for extracting machine-understandable information from natural language queries in electronic health records (EHRs).
• Traditional SP methods rely on labor-intensive, hand-built lexicons.
• Recent advancements in neural networks offer potential for more automated and flexible SP.

Purpose of the Study:

• To systematically evaluate the performance of two neural SP models for EHR question answering (QA).
• To assess the ease of application and generalizability of these advanced neural models in a clinical context.

Main Methods:

• Systematic performance assessment of two neural SP models.
• Evaluation on two distinct clinical SP datasets.
• Error analysis to identify model limitations and guide future research.

Main Results:

• The evaluated neural SP models demonstrate promising performance on clinical SP datasets.
• These models show ease of application and good generalizability for EHR QA tasks.
• Error analysis identified common failure modes of the neural models.

Conclusions:

• Advanced neural SP models offer a viable and efficient alternative to traditional methods for EHR QA.
• The findings suggest that neural SP models are a promising direction for improving clinical information retrieval.
• Further research informed by error analysis can enhance the accuracy and robustness of neural SP for clinical applications.