Presentation
of Predictive Models
The presentation
of a predictive model is important if we aim for practical application
of the model in clinical situations. Presentation of a regression
formula, or a table with regression coefficients, may be useful
to obtain insight into the prognostic effect of the predictors
included in the model, but fails to readily provide the clinician
with a predicted probability
(Harrell
et al., 1996).
Predicted
probabilities can more easily be derived from a regression model
when presented in another format, as detailed below.
Table
7.2: Presentation Formats
|
|---|
Score
Chart
| Scores
based on regression coefficients (multiplication by
10 and rounding). Scores corresponding to patient’s
predictor values added in a sumscore.
For linear models, sumscore is predicted value. For
generalized linear models, the sumscore is similar
to the linear predictor. For logistic regression models,
the sumscore is translated back to a probability through
the logistic transformation.
Survival probabilities at a certain follow-up time
can be estimated based on a predictive survival model.
| Click
for Example
Click
for Example
|
|---|
Nomogram
| The
effects of predictors are shown graphically and the
predicted value can be read directly. | Click
for Example |
|---|
Table
| Sometimes,
a table can be created which shows the predictions
according to all combinations of the values of the
predictors. When only categorical predictors are considered,
the original model can be shown. When continuous predictors
are considered, these need to be
categorized for presentation in a table. | Click
for Example |
|---|
Metamodel
| A
recent approach is to develop a model that predicts
the predictions from the original model (Harrell
et al., 1998). Such a meta-model can be simpler
in structure
than the original model, while predictive performance
is not hampered much. | Click
for Example |
|---|
|