Clinical Prediction Models
Summary
Related Notes
Model development
Development step by step
- Define aims, create a team, review literature, start writing a protocol
- Develop a new model, or update an existing one
- Define the outcome measure
- Identify candidate predictors and specify measurement methods
- Collect and examine data
- Consider sample size
- Deal with missing data
- Fit the prediction models
- Assess the performance of prediction model
- Internal validation
- Εxternal validation
- Decide on the final model
- Perform a decision curve analysis
Decision curve analysis (DCA)
DCA measures something called Net Benefit (= weighting the consequences of false positives and false negatives against patient and policy-maker preferences).
- It is a method for evaluating whether a prediction model is actually useful for making clinical decisions.
- It's useful because for healthcare the cost of a false positive (e.g., an unnecessary, invasive procedure) is usually much different than a false negative (e.g., missing a severe diagnosis).
- It basically compares predictive models against two universal baseline scenario: take action for all patients vs. take action for no patients.
- Thus, the net benefit is compared between three strategies. For example, the strategies to be compared can be:
- refer a patient for ECG if the model predicted risk exceeds 10%
- refer everyone
- refer no one
- Thus, the net benefit is compared between three strategies. For example, the strategies to be compared can be:
- Assess the predictive ability of individual predictors (optional step)
- Write up and publish