A cross-validation statistical framework for asymmetric data integration.

Area of Science:

• Biostatistics
• Bioinformatics
• Data Science

Background:

• Biobanks and public clinical datasets offer valuable resources for research.
• Integrating diverse datasets presents challenges due to heterogeneity and opaque protocols.
• Naive data integration can introduce bias and lead to unreliable conclusions.

Purpose of the Study:

• To develop a novel weighted data integration method for combining local and external clinical datasets.
• To address limitations of existing methods, including subjective weight specification and computational intractability.
• To improve parameter estimation and model prediction accuracy in the presence of data heterogeneity.

Main Methods:

• Proposed a weighted integration method minimizing local data leave-one-out cross-validation (LOOCV) error.
• Rewrote LOOCV error optimization as a function of external data integration weights for linear and Cox proportional hazards models.
• Validated the method using simulations mimicking clinical data heterogeneity and a real-world kidney transplant patient dataset.

Main Results:

• Demonstrated significant reductions in estimation error compared to existing methods.
• Showcased significant improvements in prediction error.
• Successfully applied the method to a real-world dataset from the Scientific Registry of Transplant Recipients.

Conclusions:

• The proposed weighted integration method offers an objective and computationally tractable approach for combining diverse clinical datasets.
• This method effectively reduces estimation and prediction errors, leading to more reliable model outcomes.
• The approach holds promise for enhancing biobank and public clinical data utilization in medical research.