Interpretable recurrent neural network models for dynamic prediction of the extubation failure risk in patients with invasive mechanical ventilation in the intensive care unit

Table 3 Predictive performance of the models on the test set

	Accuracy [95%CI]	F1 score [95%CI]	AUPRC [95%CI]	AUROC [95%CI]
LSTM	0.787 [0.766–0.806]	0.599 [0.585–0.612]	0.720 [0.714–0.726]	0.828 [0.809–0.846]
light_LSTM^a	0.795 [0.790–0.800]	0.603 [0.585–0.621]	0.725 [0.718–0.732]	0.827 [0.809–0.845]
GRU	0.791 [0.777–0.804]	0.599 [0.560–0.639]	0.723 [0.713–0.733]	0.829 [0.810–0.847]
light_GRU^a	0.791 [0.783–0.798]	0.585 [0.564–0.605]	0.724 [0.714–0.734]	0.825 [0.806–0.843]
Lasso LR	0.793 [0.789–0.798]	0.585 [0.573–0.597]	0.716 [0.709–0.722]	0.814 [0.795–0.832]
SVM	0.794 [0.788–0.801]	0.549 [0.533–0.566]	0.717 [0.711–0.722]	0.816 [0.797–0.834]
MLP	0.784 [0.776–0.792]	0.580 [0.547–0.613]	0.693 [0.682–0.705]	0.812 [0.792–0.830]
RF	0.790 [0.786–0.795]	0.514 [0.494–0.535]	0.726 [0.722–0.729]	0.820 [0.801–0.838]
XGB	0.790 [0.788–0.792]	0.595 [0.585–0.604]	0.724 [0.715–0.732]	0.823 [0.804–0.841]

Abbreviations: CI confidence interval, GRU gated recurrent unit, LSTM long short-term memory, LR logistic regression, MLP multi-layer perceptron, RF random forest, SVM support vector machine, XGB extreme gradient boosting
^aThe light-version RNN models based on the 26 features selected by the stepwise logistic regression model

ISSN: 1756-0381