Monday, November 30 | 9:30 a.m. to 10:30 a.m. | M3-SSBR03-2 | Room S406A
In this presentation, researchers will compare the Mirai breast imaging AI model with the Breast Cancer Surveillance Consortium (BCSC) version 2 clinical model at a diverse urban safety net hospital.
Clare Poynton, PhD, MD, from Boston University in Massachusetts will share her team’s study, which reviewed screening mammograms acquired from 2010 to 2022 and defined high-risk cases as exams followed by a biopsy-proven breast cancer within five years.
The researchers excluded patients with prior breast cancer or implants, screen-detected cancers diagnosed within six months, incomplete imaging, and missing follow-up. After excluding 399 exams for missing clinical data, the analysis covered 9,508 exams, including 727 high-risk exams. The researchers also analyzed results by race and ethnicity.
The team reports that specificity was similar for the two models (85.5% for Mirai vs. 85.3% for BCSC, p = 0.77). Mirai performed better on several other measures: Area under the receiver operating characteristic (AUROC, 0.74 vs. 0.67), sensitivity (43.3% vs. 29.6%), positive predictive value (19.8% vs. 14.3%), and negative predictive value (94.8% vs. 93.6%). In Black women, Mirai had higher positive predictive value (20.6% vs. 12.7%) and negative predictive value (95.2% vs. 93.6%). In Hispanic patients, it had higher negative predictive value (94.5% vs. 92.6%). Other differences in predictive values between groups were not statistically significant.
The authors highlight that Mirai showed greater diagnostic accuracy and predictive value than clinical risk assessment in this diverse population.
Attend this session to learn more about the model’s performance.
