A new approach to categorising continuous variables in prediction models: Proposal and validation.

Area of Science:

• Biostatistics
• Clinical Prediction Modeling
• Medical Informatics

Background:

• Clinical practitioners often categorize continuous variables for prediction models, despite statistical drawbacks like information loss.
• A valid categorization method is needed when researchers deem it necessary for clinical prediction model development.
• This practice, though statistically suboptimal, is prevalent in medical research.

Purpose of the Study:

• To propose a valid method for categorizing continuous predictors in logistic regression models.
• To optimize the discriminative ability of prediction models by maximizing the Area Under the Curve (AUC).
• To provide a practical approach for researchers who require variable categorization.

Main Methods:

• Focuses on categorizing a continuous predictor within a logistic regression framework.
• Aims to achieve the highest possible Area Under the Receiver Operating Characteristic Curve (AUC) for optimal discrimination.
• Methodology validated against known optimal cut-point locations and applied to a real-world dataset (IRYSS-COPD study).

Main Results:

• The proposed methodology provides a valid approach to categorizing continuous variables when necessary for prediction models.
• Demonstrated effectiveness in enhancing model discriminative ability, as measured by AUC.
• Successfully applied to categorize PCO2 in a real-world dataset of COPD exacerbation patients.

Conclusions:

• The study offers a statistically sound method for categorizing continuous predictors to maximize model performance (AUC).
• This approach addresses the practical need for variable categorization in clinical prediction model development.
• The method is validated and applicable to real clinical datasets, such as the IRYSS-COPD study.