Development and validation of clinical prediction models: marginal differences between logistic regression, penalized...

Area of Science:

• Medical Informatics
• Clinical Prediction Modeling
• Biostatistics

Background:

• Multivariable logistic regression is a common method for developing clinical prediction models.
• Alternative methods for model development exist but are less frequently compared.

Purpose of the Study:

• To compare the predictive accuracy of deep venous thrombosis (DVT) models developed using four distinct methods.
• To evaluate the impact of different modeling strategies on model performance.

Main Methods:

• A cohort of 2,086 primary care patients suspected of DVT was used.
• Models were developed using logistic regression, shrinkage with bootstrapping, penalized maximum likelihood estimation, and genetic programming.
• Model accuracy was assessed using discrimination and calibration in validation and cross-validation sets.

Main Results:

• Marginal differences in discrimination and calibration were observed across the four modeling methods.
• Confidence intervals for accuracy measures largely overlapped between the models.

Conclusions:

• The choice of development method had minimal impact on the predictive accuracy of DVT models.
• Effective prediction models rely more on sound modeling strategies than on methodological complexity.