Table 3:

Performance of deep learning models to detect race from chest x-rays

	Area under the receiver operating characteristics curve value for race classification
	Asian (95% CI)	Black (95% CI)	White (95% CI)
Primary race detection in chest x-ray imaging
MXR Resnet34	0·986 (0·984–0·988)	0·982 (0·981–0·983)	0·981 (0·979–0·982)
CXP Resnet34	0·981 (0·979–0·983)	0·980 (0·977–0·983)	0·980 (0·978–0·981)
EMX Resnet34	0·969 (0·961–0·976)	0·992 (0·991–0·994)	0·988 (0·986–0·989)
External validation of race detection models in chest x-ray imaging
MXR Resnet34 to CXP	0·947 (0·944–0·951)	0·962 (0·957–0·966)	0·948 (0·945–0·951)
MXR Resnet34 to EMX	0·914 (0·899–0·928)	0·983 (0·981–0·985)	0·975 (0·973–0·978)
CXP Resnet34 to MXR	0·974 (0·971–0·977)	0·955 (0·952–0·957)	0·956 (0·954–0·958)
CXP Resnet34 to EMX	0·915 (0·901–0·929)	0·968 (0·965–0·971)	0·954 (0·951–0·958)
EMX Resnet34 to MXR	0·966 (0·962–0·969)	0·970 (0·968–0·972)	0·964 (0·962–0·965)
EMX Resnet34 to CXP	0·949 (0·946–0·952)	0·973 (0·970–0·977)	0·947 (0·945–0·950)
Race detection in non-chest x-ray imaging modalities: binary race detection (Black or White)
NLST	0·92 (slice; 0·910–0·918), 0·96 (study; 0·926–0·982)	··	··
NLST to EM-CT	0·80 (slice; 0·796–0·800), 0·87 (study; 0·829–0·904)	··	··
NLST to RSPECT	0·83 (slice; 0·825–0·834), 0·90 (study; 0·836–0·958)	··	··
EM-Mammo	0·78 (slice; 0·773–0·786), 0·81 (study; 0·794–0·818)	··	··
EM-CS	0·913 (0·892–0·931)	··	··
DHA	0·87 (0·752–0·894)	··	··

Values reflect the area under the receiver operating characteristics curve for each model on the test set per slice and per study (by averaging the predictions across all slices). CXP=CheXpert dataset. DHA=Digital Hand Atlas. EM-CS=Emory Cervical Spine radiograph dataset. EM-CT=Emory Chest CT dataset. EM-Mammo=Emory Mammogram dataset. EMX=Emory CXR dataset. MXR=MIMIC-CXR dataset. NLST=National Lung Cancer Screening Trial dataset. RSPECT=RSNA Pulmonary Embolism CT dataset.