Table 2.
Mortality prediction performance of selected classifiers on various reduced data sets extracted via mutual information.
| Model and data set granularity | Average values across folds (%) | ||||||||
|
|
Specificitya (95% CI) | Sensitivitya (95% CI) | Accuracya (95% CI) | AUCb (95% CI) | |||||
| Random forest | |||||||||
|
|
25-feature data set | 83.2 (80.1-86.3) | 89.1 (86.8-91.4) | 89.2 (88.0-90.4) | 96.3 (95.9-96.6) | ||||
|
|
7-feature data set | 88.3 (86.2-90.6) | 90.0 (88.3-91.6) | 90.8 (89.8-91.7) | 95.0 (94.4-95.5) | ||||
|
|
1-feature data set | 84.9 (83.3-86.2) | 76.8 (74.9-78.7) | 80.1 (79.8-81.7) | 88.1 (87.1-89.0) | ||||
| Logistic regression | |||||||||
|
|
25-feature data set | 82.9 (79.9-85.9) | 79.6 (76.0-83.2) | 83.5 (81.9-85.1) | 88.6 (87.5-89.7) | ||||
|
|
7-feature data set | 86.5 (83.9-89.1) | 70.3 (65.4-75.2) | 79.2 (76.9-81.5) | 88.4 (87.4-89.4) | ||||
|
|
1-feature data set | 80.3 (79.4-81.3) | 80.7 (79.3-82.0) | 81.5 (81.0-82.0) | 84.2 (83.4-84.9) | ||||
aReported performance metrics represent averages across multiple simulations of 3-fold cross validation and, due to class balance variation between folds, accuracy metrics are not always a weighted average of their sensitivity and specificity.
bAUC: area under the curve obtained from the receiver operating characteristic curve.