The infusion chairs at UMass Memorial face a wall of windows, and in late afternoon the light turns the IV bags the color of iced tea. A retired school bus driver named Nancy waited for her nurse to swap lines and tapped a folded handout in her purse. “They say the scan last year looked clean,” she said, “but the computer thinks I light up next year.” She shrugged in that Massachusetts way that means both terror and practicality. “If it sees something before my doctor does, that’s okay by me.” Her hospital has been piloting an MIT–Mass General risk model that—across seven health systems from Boston to Brazil—predicted who would develop breast cancer years before tumors appeared, with concordance indices in the 0.75–0.84 range.¹
“Screening by prophecy is a different kind of medicine,” the nurse said, smiling without committing herself.
Two towns east, on Vassar Street in Cambridge, is the person who helped bend that prophecy into code. In 2014, computer scientist Regina Barzilay—already famous for teaching machines to read language—was diagnosed with breast cancer. With partners at Mass General, she trained Mirai not on neat, human-named features but on fate: here are mammograms from women who later developed cancer; here are those who didn’t; learn the difference. The tool doesn’t diagnose; it forecasts—one to five years out. Its performance held across continents; the idea is simpler still: move from firefighting to fire prevention.¹²
“The tissue itself imprints a lot of information.” Barzilay likes to say it that way—what looks like background to a radiologist can be tomorrow’s terrain to a model.³
The same logic leapt from breasts to lungs. A sibling system, Sybil, reads a single low-dose CT and forecasts lung-cancer risk for up to six years—useful precisely because many scans show nothing yet.⁴
On the other side of the country, a different alarm is getting earlier.