Table 3.
Evaluation of prediction models on the test set, after fine-tuning cut-off values for continuous variables. The 95% CIs were generated from 100 bootstrap samples of the test set
| Number of variables | mAUC (95% CI) | Generalized c-index (95% CI) | |
|---|---|---|---|
| AutoScore-Ordinal Model 1a | 8 | 0.758 (0.754, 0.762) | 0.737 (0.734, 0.741) |
| POM1a | 8 | 0.750 (0.747, 0.754) | 0.726 (0.722, 0.729) |
| RF1a | 8 | 0.767 (0.764, 0.771) | 0.547 (0.544, 0.549) |
| AutoScore-Ordinal Model 2b | 8 | 0.793 (0.789, 0.796) | 0.760 (0.757, 0.763) |
| POM2b | 8 | 0.790 (0.786, 0.793) | 0.754(0.750, 0.756) |
| RF2b | 8 | 0.798 (0.794, 0.801) | 0.564 (0.561, 0.566) |
| POM (stepwise) | 35 | 0.815 (0.812–0.819) | 0.775 (0.772–0.778) |
| POM (LASSO) | 10 | 0.704 (0.700–0.708) | 0.669 (0.665–0.673) |
POM proportional odds model, RF random forest, mAUC mean area under the curve
aThese models used the same 8 variables: emergency department (ED) length of stay (LOS), creatinine, ED boarding time, number of visits in the previous year, age, systolic blood pressure (SBP), bicarbonate and pulse
bThese models used the same 8 variables: ED LOS, creatinine, number of visits in the previous year, age, SBP, bicarbonate, pulse and metastatic cancer