Flexible co-data learning for high-dimensional prediction.

Area of Science:

• Statistical modeling
• Genomics
• Bioinformatics

Background:

• Clinical prediction with high-dimensional data is challenging.
• Complementary data (co-data) can improve prediction accuracy.
• Existing group adaptive methods have limitations with complex group structures.

Purpose of the Study:

• To develop a flexible statistical framework for clinical prediction using multiple co-data sources.
• To improve prediction performance and variable selection stability in high-dimensional settings.
• To address limitations of existing group adaptive methods.

Main Methods:

• Utilized multiple co-data sources to define covariate groups.
• Estimated adaptive multi-group ridge penalties for generalized linear and Cox models.
• Combined empirical Bayes estimation with flexible shrinkage for hyperparameter estimation.

Main Results:

• Demonstrated improved performance in cancer genomics applications.
• Showcased enhanced variable selection stability compared to other models.
• Validated the model's effectiveness across three distinct cancer genomics datasets.

Conclusions:

• The proposed method offers a versatile framework for integrating diverse co-data in clinical prediction.
• The approach effectively handles complex group structures and improves model performance.
• This method holds significant potential for advancing clinical prediction in genomics and other high-dimensional research areas.