Table 5.
Cluster outcome prediction performance.
| Cluster | Positive cases (%) | Brier loss | Accuracy | AUROC | AUPRC | Balanced accuracy | F1 | Sensitivity | Specificity |
|---|---|---|---|---|---|---|---|---|---|
| (a) Mortality | |||||||||
| −1 | 4.38 | 0.136 | 0.807 | 0.835 | 0.256 | 0.731 | 0.229 | 0.647 | 0.815 |
| 1 | 1.78 | 0.132 | 0.814 | 0.932 | 0.155 | 0.887 | 0.156 | 0.963 | 0.812 |
| 2 | 22.68 | 0.222 | 0.641 | 0.701 | 0.414 | 0.649 | 0.455 | 0.663 | 0.635 |
| 3 | 2.63 | 0.145 | 0.787 | 0.885 | 0.213 | 0.801 | 0.167 | 0.816 | 0.786 |
| 4 | 16.38 | 0.343 | 0.583 | 0.325 | 0.120 | 0.550 | 0.286 | 0.500 | 0.600 |
| 5 | 4.38 | 0.190 | 0.714 | 0.687 | 0.088 | 0.624 | 0.140 | 0.526 | 0.723 |
| 6 | 5.44 | 0.188 | 0.667 | 0.482 | 0.062 | 0.357 | 0.000 | 0.000 | 0.714 |
| 7 | 2.77 | 0.147 | 0.792 | 0.865 | 0.214 | 0.790 | 0.174 | 0.788 | 0.792 |
| All | 2.91 | 0.138 | 0.804 | 0.886 | 0.195 | 0.813 | 0.196 | 0.822 | 0.803 |
| NEWS2 | – | 0.033 | 0.925 | 0.739 | 0.127 | 0.626 | 0.201 | 0.307 | 0.945 |
| (b) Admission to higher care | |||||||||
| −1 | 11.22 | 0.180 | 0.781 | 0.879 | 0.446 | 0.806 | 0.462 | 0.837 | 0.774 |
| 0 | 1.08 | 0.217 | 0.642 | 0.785 | 0.041 | 0.713 | 0.045 | 0.786 | 0.641 |
| 1 | 4.44 | 0.200 | 0.691 | 0.754 | 0.174 | 0.682 | 0.161 | 0.672 | 0.692 |
| 2 | 37.25 | 0.210 | 0.668 | 0.737 | 0.563 | 0.672 | 0.606 | 0.686 | 0.658 |
| 3 | 7.56 | 0.221 | 0.645 | 0.687 | 0.158 | 0.625 | 0.204 | 0.603 | 0.648 |
| 4 | 40.52 | 0.279 | 0.458 | 0.421 | 0.434 | 0.436 | 0.316 | 0.300 | 0.571 |
| 5 | 7.45 | 0.197 | 0.693 | 0.755 | 0.224 | 0.662 | 0.233 | 0.625 | 0.698 |
| 6 | 11.56 | 0.263 | 0.533 | 0.667 | 0.695 | 0.593 | 0.222 | 0.667 | 0.519 |
| 7 | 14.49 | 0.224 | 0.638 | 0.692 | 0.284 | 0.635 | 0.335 | 0.630 | 0.639 |
| All | 10.23 | 0.209 | 0.670 | 0.739 | 0.246 | 0.669 | 0.293 | 0.668 | 0.670 |
| NEWS2 | – | 0.106 | 0.426 | 0.536 | 0.131 | 0.516 | 0.199 | 0.633 | 0.400 |
NEWS: National Early Warning Score; AUROC: area under receiver operating characteristic curve; AUPRC: area under precision recall curve.
Performance metrics in predicting (a) mortality and (b) admission to higher care (general ICU, cardiac ICU, and critical care unit) by implementing a logistic regression model for each cluster and the unclustered “All” patient cohort. Models are compared against predictive performance of the existing NEWS2 risk scoring system by thresholding at NEWS2 ⩾ 5 for mortality and NEWS2 ⩾ 2 for admission to higher care. Mortality was not predicted for cluster 0 since only one positive case occurred. Models with better performance than NEWS2 are highlighted in bold.