Topic Modeling for Interpretable Text Classification From EHRs.

Area of Science:

• * Computational linguistics and natural language processing applied to healthcare.
• * Machine learning for predictive modeling in clinical settings.

Background:

• * Electronic health records (EHRs) contain valuable clinical notes for predictive text classification.
• * Model interpretability is crucial for clinical decision-making.
• * Topic models offer a path to interpretable text classification by using topics as features.

Purpose of the Study:

• * To propose considerations for selecting effective topic models for EHR text classification.
• * To evaluate topic models based on both predictive performance and interpretability.
• * To identify optimal topic models for clinical predictive tasks.

Main Methods:

• * Compared 17 different topic models for text classification.
• * Evaluated models on an inpatient violence prediction task using clinical notes.
• * Assessed models using measures of interpretability and predictive performance.

Main Results:

• * No significant correlation was found between model interpretability and predictive performance.
• * The proposed fuzzy topic modeling algorithm (FLSA-W) demonstrated superior interpretability.
• * State-of-the-art methods ProdLDA and LSI achieved the highest predictive performance.

Conclusions:

• * Topic model selection for clinical text classification requires balancing interpretability and predictive accuracy.
• * FLSA-W offers strong interpretability, while ProdLDA and LSI provide robust predictive power.
• * No single topic model excels in both interpretability and predictive performance for clinical notes.