Different Forms of Clinical Prediction Models

Searching 101MeSH BrowserOnline Statistical Text'sClinical Calculator links
 


Mathematical Techniques

All clinical prediction models utilize multivariate techniques to select a limited number of variables which predict the outcome of interest.  Most models familiar to clinicians take the form of Logistic scoring systems, Recursive partitioning schemes, or Regression equations. 

 
Logistic Scoring Systems
The strength of association between each predictor and the outcome is converted to an integer value.  The total score corresponds with a probability range and/or with a specific recommendation, as in the following model to predict positive throat cultures for Streptococcus. 


 


 

Recursive partitioning

Recursive partitioning is used to develop many clinical prediction tools.   In the example of a recursive partitioning schema below,  1077 patients comprise the entire population.   Patients not fulfilling condition "A"( > 10 min since CPR started), are then removed from the model.  This leaves 845 patients.  Condition "A" was used first because it had the highest univariate association with the outcome (survival).  Condition "B" had the next highest association with survival.  If you remove patients who did not have VT or VF as their initial rhythm, and then those whose arrest was not witnessed ("C"), you are left with 119 patients.  This predicts a subset with 0 probability of survival.


 
 

Recursive partitioning models also serve as a method for clinicians to apply the model, by simply following along the flow diagram.

Regression Equations
In models usable as regression equations, values for individual predictors are plugged into an equation which was derived by multiple regression techniques. The following is an example of a regression equation for predicting MI mortality


Where B0 represents a constant, B1 represents a coefficient assigned to each predictor variable, and X1 is the value of each predictor variable (age, systolic BP, Squared BP, T wave inversions, Q waves, Initial pulse rate)

This type of model, unlike those described above, allows for input of continuous variables,  and produces a value of P from 0 to 1, unlike those models above which provide probability ranges.  However, this type of prediction tool generally require a computer or pre-programmed calculator to use.