Incorporating External Information in Tissue Subtyping: A Topic Modeling Approach.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Probabilistic topic models are used for subtype discovery but often lose disease severity information.
• Integrating discriminative loss terms with generative models is challenging due to balancing competing objectives.
• Existing methods struggle to effectively incorporate auxiliary information for improved disease severity prediction.

Purpose of the Study:

• To develop a novel framework for probabilistic topic models that incorporates external covariates into the approximate posterior.
• To improve the prediction of disease severity by leveraging discriminative power from covariates.
• To identify biologically relevant subtypes of Chronic Obstructive Pulmonary Disease (COPD) using lung CT imaging data.

Main Methods:

• Developed a framework to incorporate external covariates into the generative model's approximate posterior.
• Utilized a variant of topic model as the generative model for subtype identification.
• Applied the method to a large-scale lung CT study of COPD, using features from a neural network as covariates.

Main Results:

• The proposed method effectively integrates external covariates, improving the predictive performance for disease severity.
• Identified COPD subtypes demonstrated competitive or superior performance compared to baseline methods.
• Discovered subtypes showed correlations with genetic measurements, suggesting potential etiological relevance.

Conclusions:

• The framework offers an effective alternative for balancing generative and discriminative objectives in topic modeling.
• Incorporating discriminative covariates into the approximate posterior enhances the utility of topic models for disease subtyping and severity prediction.
• The identified COPD subtypes may represent distinct etiological pathways, warranting further investigation.