Table 2.

Performance metrics of recurrent neural network (RNN) and physicians on a balanced test set.

	Threshold-independent metrics, (95% CI)			Metrics based on a threshold of 0.5 for positive/negative classification, (95% CI)
	AUC	PR_AUC	Brier	Acc	Sens	Spec	F1	FPR	NPV	PPV
RNN	0.901 (0.870–0.932)	0.907 (0.877–0.937)	0.122 (0.088–0.156)	0.846 (0.808–0.884)	0.851 (0.798–0.904)	0.840 (0.787–0.894)	0.847 (0.797–0.897)	0.160 (0.106–0.214)	0.850 (0.797–0.903)	0.842 (0.788–0.896)
Physicians	0.745 (0.699–0.791)	0.747 (0.701–0.793)	0.217 (0.174–0.260)	0.711 (0.664–0.759)	0.594 (0.521–0.667)	0.829 (0.773–0.884)	0.673 (0.609–0.738)	0.171 (0.116–0.227)	0.671 (0.601–0.741)	0.776 (0.715–0.838)

n = 350 admissions/patients.

AUC area under curve, PR_AUC precision-recall AUC, Brier Brier score, Acc accuracy, Sens sensitivity, Spec specificity, F1 F₁-score, FPR false-positive rate, NPV negative predictive value, PPV positive predictive value, CI confidence interval.