AI-Based Breast Cancer Risk Model: Multi-institution Validation

Get Permission

In a study reported in the Journal of Clinical Oncology, Yala et al found that using the artificial intelligence (AI)-based Mirai model across diverse populations resulted in consistently accurate predictions of breast cancer risk based on screening mammography.

Study Details

In the study, Mirai was used to evaluate risk of cancer over 5 years from a screening mammogram using mammograms and pathology-confirmed breast cancer outcomes from seven different sites:

  • Massachusetts General Hospital (25,855 examinations in 7,005 patients, 588 cancers)
  • Novant (14,157 exams in 5,887 patients, 235 cancers)
  • Emory (44,008 exams in 16,495 patients, 1,003 cancers)
  • Maccabi-Assuta, Israel (6,189 exams in 6,189 patients, 186 cancers)
  • Karolinska, Sweden (19,328 exams in 7,353 patients, 1,413 cancers)
  • Chang Gung Memorial Hospital, Taiwan (13,356 exams in 13,356 patients, 244 cancers)
  • Barretos, Brazil (5,900 exams in 5,900 patients, 146 cancers).

Key Findings

A total of 128,793 mammograms from 62,185 patients were collected across the seven sites, with 3,815 exams being followed by a cancer diagnosis within 5 years. C-indices for Mirai prediction of breast cancer within 5 years of a screening mammogram were: 0.75 (95% confidence interval [CI] = 0.72–0.78) in the Massachusetts General Hospital cohort; 0.75 (95% CI = 0.70–0.80) in the Novant cohort; 0.77 (95% CI = 0.75–0.79) in the Emory cohort; 0.77 (95% CI = 0.73–0.81) in the Maccabi-Assuta cohort; 0.81 (95% CI = 0.79–0.82) in the Karolinska cohort; 0.79 (95% CI = 0.76–0.83) in the Chang Gung Memorial Hospital cohort; and 0.84 (95% CI = 0.81–0.88) in the Barretos cohort.

Receiver operating characteristic area under the curve values for Mirai performance across the seven sites ranged from 0.78 to 0.90 at 1 year, 0.76 to 0.87 at 2 years, 0.76 to 0.86 at 3 years, 0.75 to 0.85 at 4 years, and 0.75 to 0.82 at 5 years. By comparison, the traditional Tyrer-Cuzick model had a 5-year area under the curve value of 0.62 in the Massachusetts General Hospital cohort.

Analysis in the Emory cohort showed that Mirai had similar sensitivities (33.9% and 40.0%) and specificities (90.7% and 91.9%) among Black and White patients.

The investigators concluded, “Mirai, a mammography-based risk model, maintained its accuracy across globally diverse test sets from seven hospitals across five countries. This is the broadest validation to date of an AI-based breast cancer model and suggests that the technology can offer broad and equitable improvements in care.”

Adam Yala, MEng, of the Jameel Clinic, Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, is the corresponding author for the Journal of Clinical Oncology article.

Disclosure: For full disclosures of the study authors, visit

The content in this post has not been reviewed by the American Society of Clinical Oncology, Inc. (ASCO®) and does not necessarily reflect the ideas and opinions of ASCO®.