Table 3. Measurement of bias and intervention effect by diagnosis in the radiograph dataset with vision transformers.

Method	Naive data collection, AUROCa (95% CI)		Balanced empirical risk minimization, AUROC (95%CI)		AEquity-guided, AUROC (95% CI)
Label	White	Black	White	Black	White	Black
Opacity	0.8126 (0.8102-0.8149)	0.8154 (0.8131-0.8178)	0.8139 (0.8119-0.8159)	0.8169 (0.8149-0.8188)	0.8126 (0.8102-0.8149)	0.8154 (0.8131-0.8178)
Effusion	0.925 (0.9242-0.9257)	0.9195 (0.917-0.9221)	0.9249 (0.9239-0.9259)	0.9192 (0.9164-0.922)	0.925 (0.9242-0.9257)	0.9195 (0.917-0.9221)
Cardiomegaly	0.8613 (0.8603-0.8624)	0.8611 (0.8596-0.8625)	0.864 (0.8625-0.8654)	0.8647 (0.8632-0.8662)	0.8633 (0.8622-0.8644)	0.8632 (0.8619-0.8645)
Atelectasis	0.8512 (0.8497-0.8527)	0.8463 (0.8431-0.8495)	0.8508 (0.8496-0.852)	0.8457 (0.8432-0.8481)	0.8426 (0.8413-0.8439)	0.8396 (0.8375-0.8418)
Pneumonia	0.7781 (0.7769-0.7792)	0.771 (0.7677-0.7743)	0.7772 (0.7754-0.7791)	0.7706 (0.7671-0.7741)	0.7782 (0.7769-0.7794)	0.7836 (0.7803-0.7869)
Edema	0.9276 (0.9263-0.9289)	0.9303 (0.9278-0.9327)	0.925 (0.9226-0.9274)	0.9286 (0.9254-0.9318)	0.9276 (0.9263-0.9289)	0.9303 (0.9278-0.9327)
Enlarged cardiomediastinum	0.6948 (0.687-0.7025)	0.6591 (0.6422-0.6761)	0.6974 (0.6898-0.7049)	0.6588 (0.6412-0.6763)	0.6948 (0.687-0.7025)	0.6591 (0.6422-0.6761)
Pneumothorax	0.818 (0.8114-0.8247)	0.7439 (0.7123-0.7755)	0.8139 (0.8063-0.8215)	0.7408 (0.7103-0.7713)	0.8075 (0.7995-0.8155)	0.7374 (0.708-0.7668)
Consolidation	0.7903 (0.7847-0.7958)	0.7372 (0.7145-0.76)	0.7897 (0.7847-0.7947)	0.7361 (0.7136-0.7587)	0.7948 (0.7896-0.7999)	0.7443 (0.7213-0.7674)

^aAUROC: area under the receiver operating characteristic curve.