Handling missing predictor values when validating and applying a prediction model to new patients.

Area of Science:

• Clinical Epidemiology
• Biostatistics

Background:

• Missing data pose significant challenges in clinical prediction model development and application.
• Existing methods for handling missing data primarily focus on model development, with limited research on their application to individual new patients.

Purpose of the Study:

• To compare existing and novel methods for handling missing data in the context of prediction for a single new individual.
• To evaluate methods that are suitable for real-world application and model validation in practice.

Main Methods:

• Comparison of methods based on observed data only, marginalization, and fully conditional specification (chained equations) for single-patient prediction.
• Internal validation of methods in a simulation study and a cohort of prophylactic implantable cardioverter defibrillator patients.
• Performance evaluation using optimism-corrected C-statistic and mean squared prediction error, compared to multiple imputation in test patients.

Main Results:

• The study evaluated the performance of different missing data handling methods in a simulation setting.
• Methods were applied to a real-world dataset of prophylactic implantable cardioverter defibrillator patients.
• Performance metrics such as C-statistic and mean squared prediction error were used for evaluation.

Conclusions:

• Methods for handling missing data in single new patients are essential for the practical application of clinical prediction models.
• The study highlights the importance of choosing validation methods that transfer to practice.
• Further research and validation of these methods are needed for robust clinical decision-making.