Robust Data Integration Method for Classification of Biomedical Data

Area of Science:

• Oncology
• Bioinformatics
• Data Science

Background:

• Accurate cancer patient classification is crucial for effective treatment and prognosis.
• Integrating diverse data types (clinical and molecular) offers potential for improved predictive models.
• Current data integration strategies have limitations in maximizing predictive power.

Purpose of the Study:

• To develop and evaluate a novel hybrid protocol for integrating clinical and molecular data for cancer patient classification.
• To assess the performance of this hybrid approach compared to existing early and late integration methods.
• To demonstrate the utility of synthetic variables derived from molecular data in enhancing predictive models.

Main Methods:

• A hybrid data integration strategy combining early and late approaches was proposed.
• Informative clinical features were augmented with classification results from molecular datasets, creating synthetic variables.
• The protocol was applied to METABRIC breast cancer and TCGA urothelial bladder carcinoma datasets, utilizing gene expression, copy number aberrations, RNA-Seq, methylation, and protein array data.
• Performance was evaluated using repeated cross-validation and compared against early integration and late integration via super learning.

Main Results:

• The hybrid method achieved comparable performance to the best super learning variants.
• The approach facilitated sensitivity analysis and recursive feature elimination, resulting in compact predictive models.
• For breast cancer, the final model included eight clinical variables and two synthetic features.
• For urothelial bladder carcinoma, a highly accurate model was built using only two clinical features and one synthetic variable.

Conclusions:

• The inclusion of synthetic variables derived from RNA expression and copy number alterations significantly improves prognostic test quality.
• The proposed hybrid data integration protocol offers a robust and efficient method for cancer patient classification.
• This approach holds promise for broader adoption in clinical practice for enhanced cancer diagnostics and prognostics.