Fig. 3: Effects of view position on AI-based racial identity prediction. | Nature Communications

Fig. 3: Effects of view position on AI-based racial identity prediction.

From: Acquisition parameters influence AI recognition of race in chest x-rays and mitigating these factors reduces underdiagnosis bias

Fig. 3

The average scores of the racial identity prediction model were computed for each view position and compared to the average scores across all views. The average scores were computed in a weighted fashion to equally weight each patient race across the test dataset (see “Methods”). The percent change in average score per race is plotted (blue bars) compared to the differences in the frequencies of views by race (orange bars). The view frequencies are also plotted as percent changes compared to the frequency across all views. Results are shown separately for the CXP and MXR datasets. PA posterior-anterior, AP anterior-posterior, Lat lateral, Std standard, Port portable. The results are derived from the following number of images for Asian, Black, and white patients respectively: CXP-PA: 2376, 1202, 11399; CXP-AP: 11780, 5952, 64211; CXP-Lat: 2589, 1369, 12624; MXR-PA: 1904, 10867, 35855; MXR-AP: 3154, 13460, 65068; MXR-Lat: 2322, 14070, 45161; MXR-Std AP: 389, 3077, 9813; MXR-Port AP: 2765, 10383, 55255. Error bars correspond to standard deviation computed via bootstrapping. Source data are provided as a Source data file.

Back to article page