Skip to main content
Annals of Intensive Care

Table 3 Models’ performance over a range of thresholds defining lung recruitment

From: Machine learning predicts lung recruitment in acute respiratory distress syndrome using single lung CT scan

 

Validation

Test

AUC

acc

Sens

Spec

AUC

acc

Sens

Spec

Outcome: Δ45-5non-aerated tissue

  > 10% (n = 142)

0.85(0.07)

0.77(0.07)

0.78(0.10)

0.76(0.10)

0.90

0.81

0.80

0.82

  > 20% (n = 76)

0.88(0.05)

0.81(0.06)

0.81(0.12)

0.81(0.08)

0.84

0.78

0.75

0.79

  > 30% (n = 32)

0.89(0.07)

0.82(0.07)

0.81(0.18)

0.82(0.08)

0.85

0.87

0.37

0.93

Outcome: Δ15-5PaO2

  > 20 mmHg (n = 124)

0.78(0.07)

0.71(0.07)

0.71(0.11)

0.73(0.12)

0.77

0.71

0.77

0.60

  > 30 mmHg (n = 88)

0.75(0.11)

0.74(0.06)

0.73(0.13)

0.75(0.11)

0.76

0.69

0.41

0.87

  > 40 mmHg (n = 63)

0.75(0.08)

0.70(0.08)

0.57(0.20)

0.75(0.08)

0.79

0.69

0.21

0.95

  1. Lung recruitability was defined both as the percent change in not aerated tissue between 5 cmH2O and 45 cmH2O (recruiters: Δ45-5non-aerated tissue > 15%) and as the change in PaO2 between 5 cmH2O and 15 cmH2O (recruiters: Δ15-5PaO2 > 24 mmHg). Performance is evaluated in terms of mean area under the receiver operating characteristic curve (AUC), accuracy (acc), sensitivity (sens) and specificity (spec). Models were trained using the overall dataset (respectively, CT5 + G5 + G15 + M5 + M15 + RPM and CT5 + G5 + M5 + M15 + RPM) and least absolute shrinkage and selection operator (LASSO) algorithm for feature selection
Back to article page