Meta-analysis and aggregation of multiple published prediction models.

Area of Science:

• Biostatistics
• Clinical Epidemiology
• Health Informatics

Background:

• Published clinical prediction models are frequently overlooked in new model development, leading to redundant efforts and potentially underperforming models.
• Existing models may not generalize well to new populations due to differing characteristics and intended uses.

Purpose of the Study:

• To propose and evaluate novel methods for aggregating previously published prediction models.
• To improve the accuracy and generalizability of clinical prediction models by incorporating existing evidence.

Main Methods:

• Introduced two aggregation approaches: model averaging and stacked regressions.
• Developed user-friendly, stand-alone models adjusted for new validation data.
• Utilized weighting to account for model performance and heterogeneity, with different rationales for combining models.

Main Results:

• Aggregation methods demonstrated improved discrimination and calibration across most scenarios in clinical and simulation studies.
• Aggregated models performed comparably to novel models developed from scratch when validation datasets were large.
• The proposed methods accommodate models with dissimilar predictors and can be applied with limited data.

Conclusions:

• Model aggregation is a valuable strategy for leveraging existing prediction models when a validation dataset is available.
• These methods enhance prediction model performance and generalizability, offering a practical alternative to developing new models.
• The techniques are flexible, applicable even with limited data and varying model specifications.