A framework for developing, implementing, and evaluating clinical prediction models in an individual participant data...

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Individual participant data meta-analysis (IPD-MA) is increasingly used for risk prediction model development.
• Heterogeneity in baseline risk across studies presents challenges for model application and validation.
• Existing IPD-MA approaches offer averaged models but lack guidance for external use.

Purpose of the Study:

• To propose strategies for developing and validating multivariable risk prediction models from IPD-MA with heterogeneity.
• To provide guidance on selecting appropriate model intercepts for new populations.
• To evaluate model generalizability, even without external validation data.

Main Methods:

• Development of multivariable logistic regression models from IPD-MA.
• Strategies for choosing a valid model intercept.
• Internal-external cross-validation for evaluating generalizability.
• Extension of the framework to count and time-to-event data.

Main Results:

• Stratified estimation enables study-specific model intercepts.
• These intercepts inform application in new populations, even unrepresented ones.
• The proposed framework enhances model performance and generalizability.

Conclusions:

• A unified framework for IPD-MA facilitates model development, implementation, and evaluation.
• Stratified estimation and focused intercept choice improve practical application.
• Internal-external validation assesses generalizability when external data is limited.