The Mirai breast cancer risk prediction model demonstrated 29.8% concordance with the Tyrer-Cuzick model, with 79.7% of discordant cases assigned a lower risk category by Mirai.
Observational (n=84)
No
Does the Mirai breast cancer risk prediction model change the assigned risk category compared to the Tyrer-Cuzick model in women undergoing breast cancer screening?
The Mirai AI model showed modest concordance with the Tyrer-Cuzick model and frequently assigned lower risk categories, suggesting potential for deescalating short-term breast cancer screening.
e22560 Background: The current standard of care for breast cancer risk assessments is the Tyrer-Cuzick (TC) model, which requires clinical variables including family history, hormone exposure and breast density. Mirai, a validated artificial intelligence program designed to determine breast cancer risk based solely on mammographic images, requires less clinical data and predicts the probability that a woman will develop breast cancer over the next five years. Our study aims to describe Mirai performance on mammograms from a sample of women that were seen for a breast cancer risk assessment in a comprehensive cancer screening clinic within a large health system. Methods: This preliminary descriptive analysis includes 84 identified women with mammograms available for Mirai analysis and an available TC lifetime risk estimate of developing breast cancer. All patients were seen between 12/2021 and 6/2023. Women with known pathogenic variants or hereditary cancer syndromes and those with prior breast cancer diagnoses, including DCIS, were excluded. Probabilities from the Mirai model were categorized as low risk ( 1.8%. TC lifetime risk estimates were categorized as low ( 40%). Mirai results were compared to TC model results and the percentage of patients who would have a change in risk level based on using Mirai instead of TC was assessed. Results: Patients had an average age of 47 years (range: 29-72 years) and 64% were White, 13% were Asian, 7% were Black/African-American, and 15% were of unknown or another race. Lifetime breast cancer risk estimated from the TC model ranged from 6.4%-56.0% with 19.0% of patients categorized as low-risk (n = 16), 71.4% as intermediate risk (n = 60), and 9.5% as high-risk (n = 8). Comparatively, 5-year risk prediction estimates from Mirai ranged from 0.97%-5.3% with 67.9% (n = 57), 17.9% (n = 15), and 14.3% (n = 12) categorized as low, intermediate, and high-risk, respectively. Overall, 29.8% of patients (n = 25) were concordant in the assigned risk category from the TC and Mirai models. Of the 59 (70.2%) patients that had a discordant risk assessment category, 47 (79.7%) were found to be at a lower risk by the Mirai model compared to their TC score. Twelve (20.3%) patients undergoing Mirai had a higher risk compared to their TC score. Conclusions: These preliminary results from patients seen for a breast cancer risk assessment suggest an overall modest concordance in assigned risk assessment categories. The Mirai model demonstrates its potential in deescalating short-term breast cancer screening in a majority of patients seen in a high-risk clinic by using a shorter-interval, more accurate risk estimate that does not rely on historical data. Larger trials are needed to prospectively validate these findings.
Yalamanchili et al. (Thu,) conducted a observational in Breast cancer risk assessment (n=84). Mirai breast cancer risk prediction model vs. Tyrer-Cuzick (TC) model was evaluated on Concordance in assigned risk category from the TC and Mirai models. The Mirai breast cancer risk prediction model demonstrated 29.8% concordance with the Tyrer-Cuzick model, with 79.7% of discordant cases assigned a lower risk category by Mirai.