Learning Representations from Medical Text for Effective Diagnoses and Knowledge Discovery.

Area of Science:

• Computational linguistics
• Medical informatics
• Artificial intelligence in healthcare

Background:

• Medical text mining aims to extract knowledge and predict events, but few models achieve both simultaneously.
• Existing models often struggle to integrate knowledge discovery with accurate diagnostic prediction from raw clinical text.

Purpose of the Study:

• To develop a novel model for simultaneous knowledge discovery and patient diagnosis prediction using raw medical text.
• To enhance the interpretability and effectiveness of representations learned from medical text data.

Main Methods:

• Proposed the Enhanced Neural Topic Model (ENTM), a novel variant of neural topic models.
• Introduced an auxiliary loss set to improve the quality of learned representations.
• Utilized representations from ENTM to train a softmax regression model for diagnosis prediction.

Main Results:

• The ENTM model demonstrated superior performance compared to state-of-the-art pretrained neural language models on two independent medical text datasets.
• Analysis of model parameters revealed explicit semantic meanings, indicating potential for discovering diagnostic knowledge.
• The softmax regression weights highlighted significant factors contributing to diagnosis prediction.

Conclusions:

• The Enhanced Neural Topic Model (ENTM) effectively predicts patient diagnoses while simultaneously discovering knowledge from medical text.
• This approach offers a promising tool for advancing medical informatics and clinical decision support systems.
• The model's interpretability facilitates understanding the underlying factors influencing diagnostic predictions.