prev next front |1 |2 |3 |4 |5 |6 |7 |8 |9 |10 |11 |12 |13 |14 |15 |16 |17 |18 |19 |20 |21 |22 |23 |24 |25 |26 |27 |28 |29 |review

Click for larger picture

Example: 30-day mortality estimated and evaluated in a sample from the GUSTO-I trial containing 785 patients, of whom 52 died (6.6%).

We note that the actual probability corresponds well to the predicted probability, both on average and over the whole range of predictions. A regression model has intercept 0 and slope 1 for the relation between actual mortality and (the logit of) predicted probability.

The discriminative ability was reasonable, as appears from the distribution of predictions (shown at the bottom), and the spread of the 10 groups according to predicted probability (triangles).

The ROC area was 0.77, which indicates a 77% probability that a patient with the outcome had a higher predicted probability than a patient without the outcome, for a random pair of patients with and without the outcome.