Subgraph augmented non-negative tensor factorization (SANTF) for modeling clinical narrative text.

Area of Science:

• Medical informatics
• Machine learning
• Natural language processing

Background:

• Automated extraction of medical knowledge from electronic medical records (EMRs) faces challenges with scalability, selection bias, and lack of interpretability in machine learning models.
• Current machine learning approaches often function as 'black boxes' for clinicians, and training data is frequently sparse.
• There is a need for unsupervised learning methods that enhance both accuracy and interpretability in modeling clinical narrative text.

Purpose of the Study:

• To develop an unsupervised learning framework for modeling clinical narrative text.
• To improve the accuracy and interpretability of patient clustering from EMRs.
• To address the limitations of existing machine learning approaches in clinical data analysis.

Main Methods:

• Introduction of a novel framework: subgraph augmented non-negative tensor factorization (SANTF).
• SANTF utilizes atomic features (words) and automatically mines higher-order features (relations) by converting text into graph representations.
• A tensor is constructed with patients, higher-order features, and atomic features; non-negative tensor factorization is applied for patient clustering and identifying feature-patient links.

Main Results:

• SANTF achieved over 10% improvement in average F-measure for patient clustering compared to non-negative matrix factorization (NMF) and k-means.
• Established baselines using NMF and k-means with varying feature configurations.
• Identified latent groups of higher-order features providing medical insights and confirmed correlation with latent atomic feature groups.

Conclusions:

• SANTF offers a significant improvement in patient clustering accuracy from clinical narrative text.
• The framework enhances interpretability by identifying meaningful higher-order feature groups.
• This approach holds promise for advancing automated medical knowledge extraction from EMRs.