Stable Breast Cancer Prognosis.

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Accurate breast cancer prognosis is crucial for effective treatment and management.
• Existing prognostic models often assume data distribution consistency, which is frequently violated due to cancer heterogeneity and varied data collection environments.
• Data distribution shifts can compromise the stability and accuracy of current breast cancer prediction models.

Purpose of the Study:

• To develop a novel method for stable breast cancer prognosis that addresses data distribution shifts.
• To improve the reliability and accuracy of prognostic predictions in the presence of heterogeneous data.

Main Methods:

• The proposed Deep Global Balancing Cox regression (DGBCox) model leverages causal inference theory.
• High-dimensional gene expression data is transformed into latent representations using a deep autoencoder neural network.
• Causality-based balancing of latent representations is performed, followed by the selection of causal latent features for prognosis.

Main Results:

• DGBCox was applied to 12 diverse breast cancer datasets.
• The model demonstrated superior performance compared to benchmark methods.
• Results indicate enhanced prediction accuracy and stability under data distribution shifts.

Conclusions:

• DGBCox offers a robust solution for stable breast cancer prognosis, particularly in scenarios with data distribution shifts.
• The method's foundation in causal inference contributes to more reliable prognostic predictions.
• DGBCox represents a significant advancement in applying machine learning to complex cancer data.