Table 2.

Algorithm performance under four training strategies.

Strategy	Accuracy (95% CI)	AUC (95% CI)
Default (w/ Clinical Variables, w/ CheXpert Pretraining, w/ Multi-Label Loss)	0.78 (0.70, 0.85)	0.83 (0.75, 0.91)
Default w/o Clinical Variables	0.61 (0.51, 0.69)	0.57 (0.46, 0.68)
Default w/o CheXpert Pre-training	0.64 (0.55, 0.72)	0.71 (0.62, 0.81)
Default trained on dataset one, validated on dataset two	0.67 (0.62, 0.72)	0.71 (0.65, 0.76)
Default trained on dataset two, validated on dataset one	0.60 (0.55, 0.66)	0.70 (0.64, 0.76)

The default uses clinical variables, pretraining on CheXpert, and the multi-label loss.