Comparative Evaluation of Machine Learning Models for Subtyping Triple-Negative Breast Cancer: A Deep Learning-Based...

Area of Science:

• Oncology
• Bioinformatics
• Artificial Intelligence

Background:

• Triple-negative breast cancer (TNBC) presents diagnostic challenges due to its aggressive nature.
• Existing predictive models for TNBC often face limitations in data integration and accuracy.

Purpose of the Study:

• To develop and validate an innovative deep learning (DL) model for enhanced TNBC subtype and prognosis prediction using multi-omics data.
• To compare the performance of multi-omics DL models against single-omics and radiomics approaches.

Main Methods:

• Collected multi-omics data (mRNA, miRNA, gene mutations, DNA methylation) and MRI images from TCGA and TCIA databases.
• Developed a DL model optimized with Bayesian optimization for multi-omics data integration.
• Compared DL models with single-omics machine learning models and MRI radiomics models.

Main Results:

• The multi-omics DL model achieved high accuracy: 98.0% (cross-validation), 97.0% (validation set), and 91.0% (external test set) for TNBC subtype prediction.
• The DL model demonstrated superior data consistency and digital processing capabilities compared to MRI radiomics, especially in transfer testing.
• Single-omics models showed lower performance than the integrated multi-omics DL model.

Conclusions:

• The developed multi-omics DL model offers significant advancements in statistical performance and transfer learning for TNBC classification and prognosis.
• The findings highlight the potential of multi-omics data and DL algorithms to improve TNBC management.
• Further optimization of radiomics models is needed for robust cross-dataset application.