AI read a mammogram to predict five years of cancer risk

Researchers at MIT built a model called Mirai that instead learned to estimate a patient's risk of developing breast cancer over each of the next five years, trained at Massachusetts General Hospital and tested at hospitals in Sweden and Taiwan as well.

Mirai outperformed the standard clinical risk calculator, the Tyrer-Cuzick model, at identifying which patients would go on to develop breast cancer within five years, and it held up consistently across different hospitals, imaging equipment, and patient populations, unlike several earlier AI risk models that had been tested at just one site.

A risk score is not a diagnosis; it informs which patients get screened more often or earlier, a decision still made by a clinician weighing other factors too. The model's consistency across institutions was notable, since risk-prediction tools often lose accuracy outside the hospital where they were built.

The mammogram was reading five years into the future.