The point isn’t magic; it’s minutes. The FDA now says more than a thousand AI-enabled devices have cleared its pathways—a wave with radiology still in the engine room.¹²
Primary care is getting its own shortcuts. In 2018, the first autonomous diagnostic AI ever cleared by the FDA—IDx-DR—let a family clinic detect diabetic retinopathy in minutes, no ophthalmologist on site.⁶
Inside the endoscopy suite, a computer-aided system boxes tiny flickers of mucosa that slip past the human eye. In a randomized trial, adenoma detection jumped from ~40% to ~55% without lengthening the exam.⁷
On the pathology bench, Paige Prostate became the first FDA-authorized AI in digital pathology, flagging coordinates on whole-slide biopsy images so the pathologist’s eye goes straight to the likeliest trouble.⁸
Reality check. When auditors cracked open a widely used sepsis-prediction model embedded in EHRs, the system missed most cases and flooded clinicians with false alarms—triage upside down.⁹
And bias isn’t a metaphor. One hospital algorithm used cost as a proxy for need; because Black patients historically receive less costly care, the math quietly offered them less help.¹⁰
There are flameouts, too. IBM’s vaunted Watson for Oncology recommended “unsafe and incorrect” treatments, forcing a retreat to humbler goals.¹¹
Still, some bets age well. In 2020, an MIT–McMaster team used deep learning to surface an antibiotic, halicin, that killed drug-resistant pathogens in mice—a hint that AI might not only read biology but write it.¹³
If you want the human end of all this, it is as small as a woman in Sussex who got the all-clear from two radiologists—and then an AI extra reader found what eyes had missed. “I just feel so lucky,” said Sheila Tooth after a quick surgery and no chemo.¹⁴
Back in Worcester, the IV bags are still amber in the late-day light. Nancy watches hers climb and empties the last of her tea. The future is not a promise; it’s a nudge. When it works, what AI buys you is time—and the chance to spend it.
Chicago-Style Numbered Bibliography
1. Yala, Adam, et al. “Multi-Institutional Validation of a Mammography-Based Breast Cancer Risk Model.” Journal of Clinical Oncology 40, no.33 (2022): 3850–60. https://doi.org/10.1200/JCO.21.01337. A large external validation showing Mirai’s 1–5-year risk prediction held across seven health systems with C-indices ~0.75–0.84.
2. MIT Jameel Clinic. “Mirai.” Accessed October 2, 2025. https://jclinic.mit.edu/mirai/. Public page reporting Mirai’s scale (2M+ mammograms;72 hospitals;22 countries) and intended clinical use.
3. National Academy of Medicine. “Can AI Predict Breast Cancer? How a Scientist’s Personal Journey Led to an AI Model.” June 12, 2025. https://nam.edu/news-and-insights/can-ai-predict-breast-cancer/. Barzilay explains Mirai’s core idea—“the tissue itself imprints a lot of information.”
4. Mikhael, Peter G., et al. “Sybil: A Validated Deep Learning Model to Predict Future Lung Cancer Risk from a Single Low-Dose Chest CT.” Journal of Clinical Oncology 41, no.13 (2023): 2458–69. https://doi.org/10.1200/JCO.22.01345. Introduces Sybil and validates 1–6-year lung-cancer risk prediction.
5. UC Davis Health. “New AI Technology Helps Physicians Quickly Identify Stroke.” February 1, 2024. https://health.ucdavis.edu/news/headlines/new-ai-technology-helps-physicians-quickly-identify-stroke/2024/02. Local rollout with Dr. Kwan Ng’s quotes on prioritization and speed.
6. U.S. Food and Drug Administration. “IDx-DR — DEN180001: Decision Summary.” 2018. https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN180001.pdf. FDA’s official De Novo review for the first autonomous diagnostic AI.
7. Repici, Alessandro, et al. “Efficacy of Real-Time Computer-Aided Detection During Colonoscopy.” Gastroenterology