Relation-Aware Shared Representation Learning for Cancer Prognosis Analysis With Auxiliary Clinical Variables and...

Area of Science:

• Computational biology
• Bioinformatics
• Medical data analysis

Background:

• Multi-modality data integration (histopathology, genomics) advances cancer prognosis.
• Challenges include exploring inter-modality relations and handling incomplete data.
• Existing methods often overlook valuable clinical variables like grade and stage.

Purpose of the Study:

• To propose a relation-aware shared representation learning method for cancer prognosis.
• To effectively utilize clinical information and incomplete multi-modality data.
• To develop a robust prognostic model by addressing key challenges in multi-modality analysis.

Main Methods:

• Learned a multi-modal shared space using dual mapping for prognostic modeling.
• Employed relational regularizers to explore feature-label and feature-feature relationships.
• Incorporated auxiliary clinical attributes and used a partial mapping strategy for incomplete data.

Main Results:

• The proposed method demonstrated superior performance on three public datasets from The Cancer Genome Atlas (TCGA).
• Relational regularizers induced discriminatory representations and enhanced sparsity.
• Integration of clinical variables and handling of incomplete data improved model accuracy and robustness.

Conclusions:

• The developed relation-aware method offers a powerful approach for cancer prognosis using multi-modality data.
• The strategy effectively addresses challenges of inter-modality relations and data incompleteness.
• This work provides a foundation for more accurate and reliable cancer outcome prediction models.